On Wed, Feb 3, 2021 at 11:39 AM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Wed, Feb 3, 2021 at 9:22 AM Randy Dunlap <rdunlap@xxxxxxxxxxxxx> wrote:
> >
> > On 2/3/21 2:57 AM, Toke Høiland-Jørgensen wrote:
> > > Randy Dunlap <rdunlap@xxxxxxxxxxxxx> writes:
> > >
> > >> Hi,
> > >>
> > >> I see this sometimes when building a kernel: (on x86_64,
> > >> with today's linux-next 20210202):
> > >>
> > >>
> > >> CONFIG_CGROUP_BPF=y
> > >> CONFIG_BPF=y
> > >> CONFIG_BPF_SYSCALL=y
> > >> CONFIG_ARCH_WANT_DEFAULT_BPF_JIT=y
> > >> CONFIG_BPF_PRELOAD=y
> > >> CONFIG_BPF_PRELOAD_UMD=m
> > >> CONFIG_HAVE_EBPF_JIT=y
> > >>
> > >>
> > >> Auto-detecting system features:
> > >> ... libelf: [ [31mOFF[m ]
> > >> ... zlib: [ [31mOFF[m ]
> > >> ... bpf: [ [31mOFF[m ]
> > >>
> > >> No libelf found
> > >> make[5]: [Makefile:287: elfdep] Error 1 (ignored)
> > >> No zlib found
> > >> make[5]: [Makefile:290: zdep] Error 1 (ignored)
> > >> BPF API too old
> > >> make[5]: [Makefile:293: bpfdep] Error 1 (ignored)
> > >>
> > >>
> > >> but pkg-config tells me:
> > >>
> > >> $ pkg-config --modversion libelf
> > >> 0.168
> > >> $ pkg-config --libs libelf
> > >> -lelf
> > >>
> > >>
> > >> Any ideas?
> > >
> > > This usually happens because there's a stale cache of the feature
> > > detection tests lying around somewhere. Look for a 'feature' directory
> > > in whatever subdir you got that error. Just removing the feature
> > > directory usually fixes this; I've fixed a couple of places where this
> > > is not picked up by 'make clean' (see, e.g., 9d9aae53b96d ("bpf/preload:
> > > Make sure Makefile cleans up after itself, and add .gitignore")) but I
> > > wouldn't be surprised if there are still some that are broken.
> >
> > Hi,
> >
> > Thanks for replying.
> >
> > I removed the feature subdir and still got this build error, so I
> > removed everything in BUILDDIR/kernel/bpf/preload and rebuilt --
> > and still got the same libelf build error.
>
> I hate the complexity of feature detection framework to the point that
> I'm willing to rip it out from libbpf's Makefile completely. I just
> spent an hour trying to understand what's going on in a very similar
> situation. Extremely frustrating.
>
> In your case, it might be feature detection triggered from
> resolve_btfids, so try removing
> $(OUTPUT)/tools/bpf/resolve_btfids/{feature/,FEATURE-DUMP.libbpf}.
>
> It seems like we don't do proper cleanup in resolve_btfids (it should
> probably call libbpf's clean as well). And it's beyond me why `make -C
> tools/build/feature clean` doesn't clean up FEATURE-DUMP.<use-case>
> file as well.
So resolve_btfids does call libbpf's clean, but Linux Kbuild never
calls resolve_btfids' clean. Jiri, do you think that could be
improved? Basically, if something goes wrong with feature detection,
no amount of `make clean` would help and users will be forced to
struggle with frustrating experience trying to understand what's going
on.
I also still think that FEATURE-DUMP should be cleaned up by feature
infra on clean and that's not happening today, but I'm unwilling to go
and untangle all that complexity right now.
>
> >
> >
> > --
> > ~Randy
> >
