Am 07.11.2019 um 23:52 schrieb shuah shuah@kernel.org:
On 10/24/19 6:13 AM, Ilya Leoshkevich wrote:
Hello, Is there anything blocking this from getting merged?
Please see below:
This patch series fixes the following problem: linux# make kselftest TARGETS=bpf O=/mnt/linux-build # selftests: bpf: test_libbpf.sh # ./test_libbpf.sh: line 23: ./test_libbpf_open: No such file or directory # test_libbpf: failed at file test_l4lb.o # selftests: test_libbpf [FAILED] Patch 1 appends / to $(OUTPUT) in order to make it more uniform with the rest of the tree.
It isn't clear what this fixes. In addition, this patch appends an extra "/"
Excerpts from make kselftest-all with this patch applied:
linux_5.4/tools/testing/selftests/bpf//test_tcpnotify_user
I am not seeing any reason to take this patch.
The goal of patch 1 was not to fix a practical problem, but rather make kselftest code follow the same convention as the rest of the build code in tools/, namely: that $(OUTPUT) is assumed to always end with /. However, if this commonality is not valuable, I don't think I should pursue this. So I'd like to withdraw patch 1.
Patch 2 fixes the problem by prepending $(OUTPUT) to all members of $(TEST_PROGS). v1->v2:
- Append / to $(OUTPUT).
- Use $(addprefix) instead of $(foreach).
I can't reproduce this problem - I ran
make kselftest O=/tmp/linux-build
May I ask you, how did you check the test results? I'm asking because some test failures, including this one, may not affect make exit code. So you need to grep for e.g. test_libbpf.sh in make output to see the problem. For example, in my setup a lot of eBPF tests fail without this patch, and yet:
[root@bpf linux]# make kselftest TARGETS=bpf O=/mnt/linux-build <snip> ok 28 selftests: bpf: test_kmod.sh # selftests: bpf: test_libbpf.sh # ./test_libbpf.sh: line 23: ./test_libbpf_open: No such file or directory # test_libbpf: failed at file test_l4lb.o # selftests: test_libbpf [FAILED] not ok 29 selftests: bpf: test_libbpf.sh # exit=127 <snip> make[1]: Leaving directory '/mnt/linux-build'
[root@bpf linux]# echo $? 0
This might be fixing the problem in bpf case. However, I don't think this is the right fix.
It is weird in general that the existing code prepends $(OUTPUT) only to the first test prog. The first test prog is by no means special. The commit message of patch 2 explains in more detail how this discrepancy causes a practical problem:
The current logic prepends $(OUTPUT) only to the first member of $(TEST_PROGS). After that, run_one() does
cd `dirname $TEST`
For all tests except the first one, `dirname $TEST` is ., which means they cannot access the files generated in $(OUTPUT).
Is there a problem with this reasoning?
I would like to see this problem reproduced on another test first.
Best regards, Ilya