Re: linux-next: build failure after merge of the perf tree

[Ian Rogers]

On Thu, Mar 23, 2023 at 6:27 AM Arnaldo Carvalho de Melo
<acme@xxxxxxxxxx> wrote:
>
> Em Thu, Mar 23, 2023 at 09:54:37AM +1100, Stephen Rothwell escreveu:
> > Hi Ian,
> >
> > On Wed, 22 Mar 2023 11:37:31 -0700 Ian Rogers <irogers@xxxxxxxxxx> wrote:
> > >
> > > On Tue, Mar 21, 2023 at 2:40 PM Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > Hi all,
> > > >
> > > > On Fri, 17 Mar 2023 09:50:25 +1100 Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx> wrote:
> > > > >
> > > > > After merging the perf tree, today's linux-next build (native perf)
> > > > > failed like this:
> > > > >
> > > > > Auto-detecting system features:
> > > > > ...                         clang-bpf-co-re: [  [32mon [m  ]
> > > > > ...                                    llvm: [  [31mOFF [m ]
> > > > > ...                                  libcap: [  [32mon [m  ]
> > > > > ...                                  libbfd: [  [32mon [m  ]
> > > > >
> > > > > make[1]: *** Deleting file '/home/sfr/next/perf/util/bpf_skel/vmlinux.h'
> > > > > libbpf: failed to find '.BTF' ELF section in /boot/vmlinux-6.0.0-5-powerpc64le
> > > > > Error: failed to load BTF from /boot/vmlinux-6.0.0-5-powerpc64le: No data available
> > > > > make[1]: *** [Makefile.perf:1075: /home/sfr/next/perf/util/bpf_skel/vmlinux.h] Error 195
> > > > > make[1]: *** Waiting for unfinished jobs....
> > > > > make: *** [Makefile.perf:236: sub-make] Error 2
> > > > > Command exited with non-zero status 2
> > > > >
> > > > > To be clear this is a native build of perf on a PPC64le host using this
> > > > > command line:
> > > > >
> > > > > make -C tools/perf -f Makefile.perf -s -O -j60 O=../perf EXTRA_CFLAGS=-Wno-psabi
> > > > >
> > > > > (I could probably remove the EXTRA_CLFAGS now that I am building with
> > > > > gcc 12.2)
> > > > >
> > > > > I don't know which commit caused this.
> > > > >
> > > > > I have used the perf tree from next-20230316 for today.
> > > >
> > > > I am still getting this build failure.
>
> > > The build failure is intentional as not having BPF skeleton support in
> > > the perf tool will remove features. I've just sent:
> > > https://lore.kernel.org/lkml/20230322183108.1380882-1-irogers@xxxxxxxxxx/
> > > Which will recommend adding NO_BPF_SKEL=1 to your build options when
> > > this failure occurs. I didn't think a features test was appropriate
> > > for this as the feature test would basically replicate the vmlinux.h
> > > generation and I didn't want to move that support through the build
> > > system.
>
> > I was trying to understand why this step fails, but from the error
> > messages, it seems to require something to be present in the distro
> > supplied kernel image?  Is there something missing from the ppc
> > build process?  Or toolchain?  Why is it looking at the installed
> > kernel and not the built kernel?  Does the perf build now depend on the
> > kernel being built first?
> >
> > I will add NO_BPF_SKEL=1 to my build from tomorrow, but surely that
> > means that we miss some perf build testing :-(
>
> Before this BUILD_BPF_SKEL wasn't on by default, so you're back testing
> as much as before.
>
> Having said that, we need to improve the warning and I processed a patch
> from Ian to that extent:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/commit/?h=tmp.perf-tools-next&id=12a83df53444165d39d3e09fcd9627e7bec7828e
>
>  $(SKEL_OUT)/vmlinux.h: $(VMLINUX_BTF) $(BPFTOOL)
>  ifeq ($(VMLINUX_H),)
> -       $(QUIET_GEN)$(BPFTOOL) btf dump file $< format c > $@
> +       $(QUIET_GEN)$(BPFTOOL) btf dump file $< format c > $@ || \
> +       (echo "Failure to generate vmlinux.h needed for the recommended BPF skeleton support." && \
> +       echo "To disable this use the build option NO_BPF_SKEL=1." && \
> +       echo "Alternatively point at a pre-generated vmlinux.h with VMLINUX_H=<path>." && \
> +       false)
>  else
>         $(Q)cp "$(VMLINUX_H)" $@
>  endif
>
>  Which improves a bit the situation.
>
> We could check if bpftool is available and if not, suggest installing
> it.
>
> If it is available, we could check if /sys/kernel/bpf/ is available, if
> not suggest using a kernel with CONFIG_DEBUG_INFO_BTF=y, as most distros
> have by now.
>
> As to the features this enables:
>
> ⬢[acme@toolbox perf-tools-next]$ ls -la tools/perf/util/bpf_skel/*.c
> -rw-r--r--. 1 acme acme 5581 Oct 17 09:07 tools/perf/util/bpf_skel/bperf_cgroup.bpf.c
> -rw-r--r--. 1 acme acme 1764 May  5  2022 tools/perf/util/bpf_skel/bperf_follower.bpf.c
> -rw-r--r--. 1 acme acme 1438 May  5  2022 tools/perf/util/bpf_skel/bperf_leader.bpf.c
> -rw-r--r--. 1 acme acme 2290 May  5  2022 tools/perf/util/bpf_skel/bpf_prog_profiler.bpf.c
> -rw-r--r--. 1 acme acme 2164 May  5  2022 tools/perf/util/bpf_skel/func_latency.bpf.c
> -rw-r--r--. 1 acme acme 9017 Aug 26  2022 tools/perf/util/bpf_skel/kwork_trace.bpf.c
> -rw-r--r--. 1 acme acme 9251 Mar 14 08:33 tools/perf/util/bpf_skel/lock_contention.bpf.c
> -rw-r--r--. 1 acme acme 6109 Feb 27 17:47 tools/perf/util/bpf_skel/off_cpu.bpf.c
> -rw-r--r--. 1 acme acme 4310 Mar 15 11:08 tools/perf/util/bpf_skel/sample_filter.bpf.c
> ⬢[acme@toolbox perf-tools-next]$
>
> For instance, take a look at these csets:
>
> ⬢[acme@toolbox perf-tools-next]$ git log --oneline tools/perf/util/bpf_skel/lock_contention.bpf.c
> d24c0144b1dde00f perf lock contention: Show per-cpu rq_lock with address
> 1811e82767dcc6eb perf lock contention: Track and show siglock with address
> 3ace2435bb93445e perf lock contention: Track and show mmap_lock with address
> 17535a33a9c1e4fb perf lock contention: Fix compiler builtin detection
> 1bece1351c653c3d perf lock contention: Support old rw_semaphore type
> 3477f079fe70b3c9 perf lock contention: Add -o/--lock-owner option
> ebab291641bed48f perf lock contention: Support filters for different aggregation
> 5e3febe7b7b99f94 perf lock contention: Support lock addr/name filtering for BPF
> 529772c4df286159 perf lock contention: Support lock type filtering for BPF
> 688d2e8de231c54e perf lock contention: Add -l/--lock-addr option
> eca949b2b4addd94 perf lock contention: Implement -t/--threads option for BPF
> fd507d3e359c7e06 perf lock contention: Add lock_data.h for common data
> c66a36af7ba3a628 perf lock contention: Do not use BPF task local storage
> 433b31fa00797a2a perf lock contention: Fix a build error on 32-bit
> c1da8dd5c11dabd5 perf lock contention: Skip stack trace from BPF
> 6d499a6b3d90277d perf lock: Print the number of lost entries for BPF
> 6fda2405f414b24a perf lock: Implement cpu and task filters for BPF
> 407b36f69efbdccf perf lock: Use BPF for lock contention analysis
> ⬢[acme@toolbox perf-tools-next]$
>
>
> or:
>
> ⬢[acme@toolbox perf-tools-next]$ git log --oneline tools/perf/util/bpf_skel/kwork_trace.bpf.c
> acfb65fe1d11a97f perf kwork: Add workqueue trace BPF support
> 5a81927a407c050a perf kwork: Add softirq trace BPF support
> 420298aefe94840f perf kwork: Add IRQ trace BPF support
> daf07d220710a3c8 perf kwork: Implement BPF trace
> ⬢[acme@toolbox perf-tools-next]$
>
> So we need to fine tune this detection of needed components to build
> these features, in time for the v6.4 merge window, that is why we
> decided  to make this opt-out to hammer out problems.
>
> Other arches probably will hit some of these problems, lets try to
> encourage others to try what is in linux-next.

Just to throw out, perhaps we could ship a vmlinux.h in the perf
source tree. We could also validate it with something similar to the
header file checking. This would remove the bpftool and properly built
kernel issues. It may make a package maintainer's life easier.

Thanks,
Ian

> - Arnaldo