Re: [PATCH v2 bpf-next 2/6] libbpf: rename static variables during linking

From: Alexei Starovoitov <alexei.starovoitov@gmail.com>
To: Andrii Nakryiko <andrii.nakryiko@gmail.com>
Cc: Yonghong Song <yhs@fb.com>, Andrii Nakryiko <andrii@kernel.org>,
	bpf <bpf@vger.kernel.org>, Networking <netdev@vger.kernel.org>,
	Alexei Starovoitov <ast@fb.com>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Kernel Team <kernel-team@fb.com>
Subject: Re: [PATCH v2 bpf-next 2/6] libbpf: rename static variables during linking
Date: Fri, 23 Apr 2021 16:48:34 -0700	[thread overview]
Message-ID: <CAADnVQKo+efxMvgrqYqVvUEgiz_GXgBVOt4ddPTw_mLuvr2HUw@mail.gmail.com> (raw)
In-Reply-To: <CAEf4BzadCR+QFy4UY8NSVFjGJF4CszhjjZ48XeeqrfX3aYTnkA@mail.gmail.com>

On Fri, Apr 23, 2021 at 4:35 PM Andrii Nakryiko
<andrii.nakryiko@gmail.com> wrote:
>
> On Fri, Apr 23, 2021 at 4:06 PM Alexei Starovoitov
> <alexei.starovoitov@gmail.com> wrote:
> >
> > On Fri, Apr 23, 2021 at 2:56 PM Yonghong Song <yhs@fb.com> wrote:
> > > >>>
> > > >>> -static volatile const __u32 print_len;
> > > >>> -static volatile const __u32 ret1;
> > > >>> +volatile const __u32 print_len = 0;
> > > >>> +volatile const __u32 ret1 = 0;
> > > >>
> > > >> I am little bit puzzled why bpf_iter_test_kern4.c is impacted. I think
> > > >> this is not in a static link test, right? The same for a few tests below.
> > > >
> > > > All the selftests are passed through a static linker, so it will
> > > > append obj_name to each static variable. So I just minimized use of
> > > > static variables to avoid too much code churn. If this variable was
> > > > static, it would have to be accessed as
> > > > skel->rodata->bpf_iter_test_kern4__print_len, for example.
> > >
> > > Okay this should be fine. selftests/bpf specific. I just feel that
> > > some people may get confused if they write/see a single program in
> > > selftest and they have to use obj_varname format and thinking this
> > > is a new standard, but actually it is due to static linking buried
> > > in Makefile. Maybe add a note in selftests/README.rst so we
> > > can point to people if there is confusion.
> >
> > I'm not sure I understand.
> > Are you saying that
> > bpftool gen object out_file.o in_file.o
> > is no longer equivalent to llvm-strip ?
> > Since during that step static vars will get their names mangled?
>
> Yes. Static vars and static maps. We don't allow (yet?) static
> entry-point BPF programs, so those don't change.
>
> > So a good chunk of code that uses skeleton right now should either
> > 1. don't do the linking step
> > or
> > 2. adjust their code to use global vars
> > or
> > 3. adjust the usage of skel.h in their corresponding user code
> >   to accommodate mangled static names?
> > Did it get it right?
>
> Yes, you are right. But so far most cases outside of selftest that
> I've seen don't use static variables (partially because they need
> pesky volatile to be visible from user-space at all), global vars are
> much nicer in that regard.

Right.
but wait...
why linker is mangling them at all and why they appear in the skeleton?
static vars without volatile should not be in a skeleton, since changing
them from user space might have no meaning on the bpf program.
The behavior of the bpf prog is unpredictable.
Only volatile static can theoretically be in the skeleton, but as you said
probably no one is using them yet, so we can omit them from skeleton too.