BPF selftests build failures
Hi BPF and selftest maintainers. I get a very strange failure when trying to build the bpf selftests on current net-next master:
hch@brick:~/work/linux/tools/testing/selftests/bpf$ make GEN vmlinux.h Error: failed to load BTF from /home/hch/work/linux/vmlinux: No such file or directory make: *** [Makefile:187: /home/hch/work/linux/tools/testing/selftests/bpf/tools/include/vmlinux.h] Error 2 make: *** Deleting file '/home/hch/work/linux/tools/testing/selftests/bpf/tools/include/vmlinux.h'
But vmlinux obviously exists:
hch@brick:~/work/linux/tools/testing/selftests/bpf$ ls -l ~/work/linux/vmlinux -rwxr-xr-x 1 hch hch 628909168 Jul 20 09:48 /home/hch/work/linux/vmlinux
.config attached.
Quoted 6 lines · Christoph Hellwig
Hi BPF and selftest maintainers. I get a very strange failure when trying to build the bpf selftests on current net-next master:
hch@brick:~/work/linux/tools/testing/selftests/bpf$ make GEN vmlinux.h Error: failed to load BTF from /home/hch/work/linux/vmlinux: No such file or directory
That's bpftool complaining that BTF is not present in vmlinux. You need CONFIG_DEBUG_INFO_BTF=y and pahole >= v1.16 You also need llvm 10 to build bpf progs.
Quoted 4 lines · Alexei Starovoitov
Quoted 3 lines · Christoph Hellwig
Hi BPF and selftest maintainers. I get a very strange failure when trying to build the bpf selftests on current net-next master:
Quoted 8 lines
Quoted 3 lines
hch@brick:~/work/linux/tools/testing/selftests/bpf$ make GEN vmlinux.h Error: failed to load BTF from /home/hch/work/linux/vmlinux: No such file or directoryThat's bpftool complaining that BTF is not present in vmlinux. You need
CONFIG_DEBUG_INFO_BTF=yand pahole >= v1.16 You also need llvm 10 to build bpf progs.
These never ending bumps to required versions to build kernels and bpf code are not friendly to users. Until a recent commit I was able to use Ubuntu 20.04 for bpf development and testing (e.g., the recent devmap changes). Ubuntu 20.04 is an LTS OS released just 3 months ago with pahole 1.15 and llvm 10. Now with v5.8-next 20.04 is out of date.
These hard requirements are making BTF inaccessible to average users by *requiring* them to constantly chase new build versions. At this point I can no longer point potential BPF/XDP users to this latest OS and say try it out. Any instructions I write for getting started will be out of date in some arbitrarily short period of time.
Worst, it is not even feature based within BTF; it is all or nothing. ie., try to use 5.8-next with Ubuntu 20.04 and you lose all of BTF for all use cases, not just some recently added feature.
If you want people to adopt this tech, you need to make it easier to use.
Quoted 11 lines · Alexei Starovoitov
Quoted 6 lines · Christoph Hellwig
Hi BPF and selftest maintainers. I get a very strange failure when trying to build the bpf selftests on current net-next master:
hch@brick:~/work/linux/tools/testing/selftests/bpf$ make GEN vmlinux.h Error: failed to load BTF from /home/hch/work/linux/vmlinux: No such file or directoryThat's bpftool complaining that BTF is not present in vmlinux. You need
CONFIG_DEBUG_INFO_BTF=yand pahole >= v1.16 You also need llvm 10 to build bpf progs.
That's not a very useful error message then. Can you fix it to tell people what is wrong?
Quoted 11 lines · Alexei Starovoitov
Quoted 6 lines · Christoph Hellwig
Hi BPF and selftest maintainers. I get a very strange failure when trying to build the bpf selftests on current net-next master:
hch@brick:~/work/linux/tools/testing/selftests/bpf$ make GEN vmlinux.h Error: failed to load BTF from /home/hch/work/linux/vmlinux: No such file or directoryThat's bpftool complaining that BTF is not present in vmlinux. You need
CONFIG_DEBUG_INFO_BTF=yand pahole >= v1.16 You also need llvm 10 to build bpf progs.
Oh well, after the non-obvious enabling of CONFIG_DEBUG_INFO_BTF I did run into the too old clang (clang 9) as well, which just creates random errors instead of warning about a too old compiler, sight.
Then I installed clang-10, but there still seem to be various failures in the net-next baseline, and the setsockopt code I'm trying to test doesn't even get exercised from sticking a printk in there.
That is not a great testcase experience..