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

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Fri, Apr 23, 2021 at 5:22 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Fri, Apr 23, 2021 at 5:13 PM Andrii Nakryiko
> <andrii.nakryiko@gmail.com> wrote:
> >
> > On Fri, Apr 23, 2021 at 4:48 PM Alexei Starovoitov
> > <alexei.starovoitov@gmail.com> wrote:
> > >
> > > 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.
> >
> > It's up to the compiler. If compiler decides that it shouldn't inline
> > all the uses (or e.g. if static variable is an array accessed with
> > index known only at runtime, or many other cases where compiler can't
> > just deduce constant value), then compiler will emit ELF symbols, will
> > allocate storage, and code will use that storage. static volatile just
> > forces the compiler to not assume anything at all.
> >
> > If the compiler does inline all the uses of static, then we won't have
> > storage allocated for it and it won't be even present in BTF. So for
> > libbpf, linker and skeleton statics are no different than globals.
>
> Not quite.
> The compiler can inline static values in some cases and still
> keep the var in global data in others places in the same C file.
> If it's a static struct the compiler is free to do any crazy optimization
> on this data it can come up with.
> So the presence of it in the elf symbols doesn't guarantee anything.

Ok, yes, I can see how that could happen. But this was always the case
when someone used statics without volatile. And I've seen recent
example of using plain static (though for internal state keeping, not
used from user-space, if I remember correctly). So we could drop
non-volatile statics from skeleton, I suppose, but I'm afraid we might
be breaking existing users.

>
> > Static maps are slightly different, because we use SEC() which marks
> > them as used, so they should always be present.
>
> yes. The used attribute makes the compiler keep the data,
> but it can still inline it and lose the reference in the .text.

At least if the map is actually used with helpers (e.g.,
bpf_map_lookup_elem(&map, ...)) it would be invalid for compiler to do
anything crazy with that map reference, because compiler has no
visibility into what opaque helpers do with that memory. So I don't
think it can alias multiple maps, for instance. So I think static maps
should be fine.

>
> > Sub-skeleton will present those statics to the BPF library without
> > name mangling, but for the final linked BPF object file we need to
> > handle statics. Definitely for maps, because static means that library
> > or library user shouldn't be able to just extern that definition and
> > update/lookup/corrupt its state. But I think for static variables it
> > should be the same. Both are visible to user-space, but invisible
> > between linked BPF compilation units.
>
> I agree with motivation and also agree that static map def is a clean
> way to hide such a map inside the library.
> I'm not sure yet that static + used is enough.
> More study needed.
> For maps it's only the reference inside .text that is relevant.
> If the compiler messes with it then we might have issues.
> Like two static maps might look the same to the compiler
> They will have different field names, but their contents are zeros
> in both cases. I'm not sure that the compiler will not decide to merge them
> into one data location. In such case .text insns will be pointing
> to the same data offset. The symbol names are probably going to be different.
> So we might be ok with static+used only.

See above about passing a pointer to map into black box functions. I'd
bet that the compiler can't merge together two different references at
least because of that.

For static maps, btw, just like for static functions and vars, there
is no symbol, it's an offset into .maps section. We use that offset to
identify the map itself.