On Mon, Feb 24, 2020 at 01:21:23PM +0000, Alexandru Elisei wrote: > Hi Naresh, > > On 2/24/20 12:53 PM, Naresh Kamboju wrote: > > [Sorry for the spam] > > > > Greeting from Linaro ! > > We are running kvm-unit-tests on our CI Continuous Integration and > > testing on x86_64 and arm64 Juno-r2. > > Linux stable branches and Linux mainline and Linux next. > > > > Few tests getting fail and skipped, we are interested in increasing the > > test coverage by adding required kernel config fragments, > > kernel command line arguments and user space tools. > > > > Your help is much appreciated. > > > > Here is the details of the LKFT kvm unit test logs, > > > > [..] > > I am going to comment on the arm64 tests. As far as I am aware, you don't need any > kernel configs to run the tests. > > From looking at the java log [1], I can point out a few things: > > - The gicv3 tests are failing because Juno has a gicv2 and the kernel refuses to > create a virtual gicv3. It's normal. Yup > > - I am not familiar with the PMU test, so I cannot help you with that. Where is the output from running the PMU test? I didn't see it in the link below. > > - Without the logs, it's hard for me to say why the micro-bench test is failing. > Can you post the logs for that particular run? They are located in > /path/to/kvm-unit-tests/logs/micro-bench.log. My guess is that it has to do with > the fact that you are using taskset to keep the tests on one CPU. Micro-bench will > use 2 VCPUs to send 2^28 IPIs which will run on the same physical CPU, and sending > and receiving them will be serialized which will incur a *lot* of overhead. I > tried the same test without taskset, and it worked. With taskset -c 0, it timed > out like in your log. We've also had "failures" of the micro-bench test when run under avocado reported. The problem was/is the assert_msg() on line 107 is firing. We could probably increase the number of tries or change the assert to a warning. Of course micro-bench isn't a "test" anyway so it can't "fail". Well, not unless one goes through the trouble of preparing expected times for each measurement for a given host and then compares new results to those expectations. Then it could fail when the results are too large (some threshold must be defined too). > > - there are also other tests that spawn multiple VCPUs, using taskset will > serialize the VCPUs and will probably hide any potential locking issues. Indeed. Thanks, drew > > [1]|https://lkft.validation.linaro.org/scheduler/job/1242488| > > |Thanks,| > |Alex| > |||| >
