On Wed, Jun 12, 2019 at 9:32 PM shuah shuah@kernel.org wrote:
On 6/12/19 12:29 PM, Dmitry Vyukov wrote:
On Wed, Jun 12, 2019 at 6:45 PM shuah shuah@kernel.org wrote:
Hi Dmitry,
This is the 6th email from you in a span of 3 hours! I am just going to respond this last one. Please try to summarize your questions instead of sending email storm, so it will be easier to parse and more productive for both of us.
Hi Shuah,
Sorry for that. Let me combine all current questions in a more structured way.
My motivation: I am trying to understand what does it take to run/add kernel tests in particular for the purpose of providing working instructions to run kernel test to a new team member or a new external kernel developer, and if it's feasible to ask a kernel developer fixing a bug to add a regression test and ensure that it works. Note in these cases a user may not have lots of specific expertise (e.g. any unsaid/implicit thing may be a showstopper) and/or don't have infinite motivation/time (may give up given a single excuse to do so) and/or don't have specific interest/expertise in the tested subsystem (e.g. a drive-by fix). So now I am trying to follow this route myself, documenting steps.
- You suggested to install a bunch of packages. That helped to some
degree. Is there a way to figure out what packages one needs to install to build the tests other than asking you?
I have to go through discovery at times when new tests get added. I consider this a part of being a open source developer figuring out dependencies for compiling and running. I don't have a magic answer for you and there is no way to make sure all dependencies will be documented.
The problem with this is that all of this needs to be figured again and again, thousands of times. And since testing is now always considered as a first-class citizen by developers, requirement to do any additional step leads to the fact that testing is not done at all, tests are not added at all, etc. If for, say, building kernel users are usually ready to jump through all possible hoops just because they don't have any other choice; testing needs to be as simple as possible to be used.
Documenting is good. But everything captured in machine usable form is even better because it allows further automation. One way to do it is to either provide a reference system, or provide a minimal system and then each test specifying all additional requirements (kernel configs, hardware, host packages, target packages, etc). Then this can be tested by installing all specified and building/running the test. If it works, anybody is able to trivially repeat the same. If it does not work, we caught broken/regressed test.
- Build of bpf tests was broken after installing all required
packages. It helped to delete some random files (tools/testing/selftests/bpf/{feature,FEATURE-DUMP.libbpf}). Is it something to fix in kselftests? Deleting random files was a chaotic action which I can't explain to anybody.
I am still getting 1 build error:
CC /usr/local/google/home/dvyukov/src/linux/tools/testing/selftests/bpf/str_error.o
timestamping.c:249:19: error: ‘SIOCGSTAMP’ undeclared (first use in this function); did you mean ‘SIOCGSTAMPNS’?
What should I do to fix this?
I am not seeing that on my system. I suspect you are still missing some packages and/or headers.
Shouldn't the test use the headers from the kernel source? Docs say:
* First use the headers inside the kernel source and/or git repo, and then the system headers. Headers for the kernel release as opposed to headers installed by the distro on the system should be the primary focus to be able to find regressions.
And it makes sense because it both avoids the problem of installing unknown set of packages and more importantly testing of things that are not yet present in headers of all distros.
- Are individual test errors are supposed to be fatal? Or I can just
ignore a single error and proceed?
Individual test errors aren't fatal and the run completes reporting errors from individual tests.
I've tried to proceed, but I am not sure if I will get some unexplainable errors later because of that. By default I would assume that any errors during make are fatal.
Kselftest is a suite of developer tests. Please read the documentation.
I read Documentation/dev-tools/kselftest.rst again but I don't see anything about this. Which part of the doc are you referring to?
The instructions on running tests:
$ make -C tools/testing/selftests run_tests $ make kselftest
Do they assume that the tests will run right on my host machine? It's not stated/explained anywhere, but I don't see how "make kselftest" can use my usual setup because it don't know about it.
You have to tailor to it your environment. This is really for kernel developers and test rings that routines test kernels under development.
It's good when a system allows for fine-tuning by experts when they are willing to do so. But it's equally important to provide the default 1-step way of using a system for people who don't yet have the expertise, or maybe not willing to invest lots of time. I think it would be positive for kernel code quality to allow new kernel developers to run kernel test as the first thing they do with kernel, rather then require years of experience first. What do you think regarding reducing the entry barrier for kernel testing?
I cannot run tests on the host. Policy rules aside, this is yet untested kernel, so by installing it I am risking losing my whole machine.
This is just like running kernel make. Build happens on the system. The idea is that kernel developers use these tests to test their code.
Reading further, "Install selftests" and "Running installed selftests" sections seem to be a way to run tests on another machine. Is it correct? Are there any other options? There seems to be a bunch of implicit unsaid things, so I am asking in case I am missing some even simpler way to run tests. Or otherwise, what is the purpose of "installing" tests?
- The "Running installed selftests" section says:
"Kselftest install as well as the Kselftest tarball provide a script named "run_kselftest.sh" to run the tests".
Right. You have to generate it. As documented in the kselftest.rst. kselftest_install.sh will install compiled tests and run_skefltest.sh
You can run gen_kselftest_tar.sh to create tarball and unpack it on your test system.
What is the "Kselftest tarball"? Where does one get one? I don't see any mentions of "tarball" anywhere else in the doc.
Please see above. You can generate tarball yourself using "tar"
I see. gen_kselftest_tar.sh may be something worth mentioning in the doc.
- What image am I supposed to use to run kselftests? Say, my goal is
either running as many tests as possible (CI scenario), or tests for a specific subsystem (a drive-by fix scenario). All images that I have do not seem to be suitable. One is failing with: ./run_kselftest.sh: 2: ./run_kselftest.sh: realpath: not found And there is no clear path to fix this. After I guessed the right package to install, it turned out to be broken in the distro. In another image all C programs fail to run with: ./test_maps: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.26'
How is one supposed to get an image suitable for running kselftests?
When you say image - what is image in this context.
You build new kernel, install it, boot the new kernel and run selftests. If your kernel build system is different from test system, then you build and install kernel and build kselftest and install kselftest and copy them over to the test system.
By image I mean disk image, what provides the user-space system, binaries, etc. kselftests seem to make a bunch of assumptions about the disk image required to run them, so I am looking for a way to create the image suitable for running kselftest.
- Lots of tests fail/skipped with some cryptic for me errors like:
# Cannot find device "ip6gre11"
# selftests: [SKIP] Could not run test without the ip xdpgeneric support
Right that means the kernel doesn't support the feature.
# modprobe: ERROR: ../libkmod/libkmod.c:586 kmod_search_moddep() could not open moddep file '/lib/modules/5.1.0+/modules.dep.bin'
You don't have the module built-in.
# selftests: bpf: test_tc_edt.sh # nc is not available not ok 40 selftests: bpf: test_tc_edt.sh
Say, I either want to run tests for a specific subsystem because I am doing a drive-by fix (a typical newcomer/good Samaritan scenario), or I want to run as many tests as possible (a typical CI scenario). Is there a way to bulk satisfy all these prerequisite (configs, binaries and whatever they are asking for)?
- There is a test somewhere in the middle that consistently reboots my machine:
# selftests: breakpoints: step_after_suspend_test [ 514.024889] PM: suspend entry (deep) [ 514.025959] PM: Syncing filesystems ... done. [ 514.051573] Freezing user space processes ... (elapsed 0.001 seconds) done. [ 514.054140] OOM killer disabled. [ 514.054764] Freezing remaining freezable tasks ... (elapsed 0.001 seconds) done. [ 514.057695] printk: Suspending console(s) (use no_console_suspend to debug) early console in extract_kernel ...
Yes. Some tests require reboot and you want to avoid those, if you don't want to run them. Please look at the kselftest.rst.
I've read it several times. I don't see anything relevant, which part are you referring to? I've searched for "reboot" and "breakpoint" specifically, but also don't see any matches.
Wouldn't it be more useful to disable this test by default? Users interested specifically in this test can enable it them. But it looks like it does not lead to a useful behavior for anybody else.
Is it a bug in the test? in the kernel? Or how is this supposed to work/what am I supposed to do with this?
- Do you know if anybody is running kselftests? Running as in
running continuously, noticing new failures, reporting these failures, keeping them green, etc. I am asking because one of the tests triggers a use-after-free and I checked it was the same 3+ months ago. And I have some vague memories of trying to run kselftests 3 or so years ago, and there was a bunch of use-after-free's as well.
Yes Linaro test rings run them and kernel developers do. I am cc'ing Naresh and Anders to help with tips on how they run tests in their environment. They have several test systems that they install tests and run tests routine on all stable releases.
Naresh and Anders! Can you share your process for running kselftest in Linaro test farm. Thanks in advance.
- Do we know what's the current code coverage achieved by kselftests?
What's covered? What's not? Overall percent/per-subsystem/etc?
No idea.
- I am asking about the aggregate result, because that's usually the
first thing anybody needs (both devs testing a change and a CI). You said that kselftest does not keep track of the aggregate result. So the intended usage is always storing all output to a file and then grepping it for "[SKIP]" and "[FAIL]". Is it correct?
As I explained, Kselftest will not give you the aggregate result.
Is it just not implemented? Or it's some kind of principal position? It's just a bit surprising to me because the aggregate result is the first thing anybody is looking for after running a set of tests...