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

On Tue, May 4, 2021 at 10:22 PM Alexei Starovoitov
[off-list ref] wrote:

On Mon, May 3, 2021 at 9:42 PM Alexei Starovoitov
[off-list ref] wrote:

quoted

On Wed, Apr 28, 2021 at 12:33:36PM -0700, Andrii Nakryiko wrote:

quoted

quoted

At least I'm only starting to grasp the complexity of the problem.

I did and didn't find anything satisfactory. But I think we are coming
at this from two different angles, which is why we can't agree on
anything. So just a reminder, static is about two properties:
    1) access protection
    2) naming collisions.

I'm trying to let name collisions on BPF side happen and be allowed
*while* also allowing access to those same name-collisioned entities
(maps and vars, both) from user-space in some non-random fashion. That
inevitably requires some compromises/conventions on the user-space
side. Such an approach preserves both 1) and 2).

You are trying to enforce unique names (or at least aliases) for
static variables, if I understand correctly, which preserves 1) at the
expense of 2). It seems to be a similar idea with custom SEC(), though
you ignored my request to elaborate on how you see that used, so I'm
guessing here a bit.

But I think we can get just 1) with global variables with custom
visibilities. E.g., just marking map/variable as __hidden would
disallow extern'ing it from other files. That's obviously limiting for
extern'ing within the library, so we can keep digging deeper and
define __internal (STV_INTERNAL) that would be "upgraded" to
STV_HIDDEN after the initial linking pass. So you'd compile your BPF
library with __internal, but your lib.bpf.o will have those global
variables as STV_HIDDEN and thus inaccessible from other libraries and
BPF app itself.

So if we are ok breaking existing static variable users, then just
dropping statics from BPF skeleton and supporting extra __hidden and
__internal semantics for variables and maps would bypass these issues.
I wanted statics mostly for property 2), but if I can't get it, then
I'd drop statics from skeletons altogether.

If I could drop statics for skeletons that were statically linked,
that wouldn't be a regression. It's impossible to do right now, but we
can also add a new SHT_NOTE section, which we can use to detect
statically linked vs Clang-generated .bpf.o. Certainly more ELF
fussing around than I'd like, but not the end of the world either.

Thoughts? Did that summarize the issue well enough?

Background for all:

Until Nov 2019 libbpf didn't support global variables, so bpf programs
contained code like 'static volatile const int var = 1;'
Then the skeleton was introduced which went through BTF of a given
datasec and emitted all variables from that section into .skel.h.
It didn't bother filtering static vs global variables, so
static vars in *.bpf.c world became visible into user space *.c world.
While libbpf supported single bpf.o file such extern-ing of statics
was fine, but with support of linking multiple *.bpf.o there
is a question of what to do with static variables with the same names
in different files.

Consider the following scenario:
One bpf developer creates a library conntrack. It has
impl.bpf.c
ct_api.bpf.c
and corresponding user space ct.c that uses skel.h to access
data in these two bpf files.

Another bpf developer creates a library for lru. It has
impl.bpf.c
lru_api.bpf.c
and corresponding user space lru.c.

Now the 3rd developer is writing its main.bpf.c and wants to use these libs.

The libs should be usable in pre-compiled form. The availability of
the source code is nice, but it shouldn't be mandatory.

So there is libct.a (with user space) and libct.bpf.a (with bpf code)
and liblru.a (user) and liblru.bpf.a (bpf code).

The developer should be able to link
main.bpf.o liblru.bpf.a libct.bpf.a
into final_main.bpf.o
And link main.o liblru.a libct.a with user space bits into a.out.

The lru.skel.h and ct.skel.h used by these libs were generated
out of corresponding *.bpf.o and independent of each other.
There should be no need to recompile user space lru.c and ct.c after
linking of final_main.bpf.o and generating final skeleton.

I think all three developers should be able to use static variables
in their .bpf.c files without worrying about conflicts across three
projects.
They can use global vars with __attribute__("hidden"),
but it's not equivalent to static. The linker will complain of
redefinition if the same name is used across multiple files
or multiple libs.
So doing 'int var __attribute__("hidden");' in libct.bpf.a and
in liblru.bpf.a will prevent linking together.
That's traditional static linking semantics.

Using file name as a prefix for static vars doesn't work in general,
since file names can be the same.
What can work is the library name. The library name is guaranteed to be
unique in the final linking phase.
I think we can use it to namespace static variables across
three sets of bpf programs.
Also I think it's ok to require a single developer to enforce
uniqueness of static vars within a project.

In other words 'static int a;' in impl.bpf.c will conflict
with 'static int a;' in ct_api.bpf.c
But the static variable in ct_api.bpf.c will not conflict
with the same variable in lru_api.bpf.c and will not conflict
with such var in main.bpf.c because they're in a different namespaces.

Here are few ways for the programmer to indicate the library namespaces:

- similar to 'char license[]' use 'char library[]="lru";' in *.bpf.c
The static linker will handle this reserved name specially just like
it does 'license' and 'version'.

- #pragma clang attribute push (__attribute__((annotate("lib=lru"))), apply_to = variable)

- #pragma comment(lib, "lru")

I think it's important to define namespaces within *.bpf.c.
Defining them outside on linker command line or linker script is cumbersome.

I think combining *.o into .a can happen with traditional 'ar'. No need for
extra checks for now.
The linking of main.bpf.o liblru.bpf.a libct.bpf.a
will fail if static vars with the same name are present within the same library.
The library namespaces will prevent name conflicts across libs and main.bpf.o
If namespace is not specified it means it's empty, so the existing
hacks of 'static volatile const int var;' will continue working.

The skeleton can have library name as anon struct in skel.h.
All vars can be prefixed too, but scoping them into single struct is cleaner.

I think it doesn't hurt if final_main.skel.h includes all bpf vars from lru and
ct libraries, but I think it's cleaner to omit them.

It's not clear to me yet how final_main__open() and final_main__load() skeleton
methods will work since lru and ct libs might need their specific initialization
that is done by user space lru.c and ct.c.
Also the whole scheme should work with upcoming light skeleton too.
The design for bpf libraries should accommodate signed libraries.

All of the above is up for discussion. I'd love to hear what golang folks
are thinking, since above proposal is C centric.

I want to clarify a few things that were brought up in offline discussions.
There are several options:
1. don't emit statics at all.
That will break some skeleton users and doesn't solve the name conflict issue.
The library authors would need to be careful and use a unique enough
prefix for all global vars (including attribute("hidden") ones).
That's no different with traditional static linking in C.
bpf static linker already rejects linking if file1.bpf.c is trying to
'extern int foo()'
when it was '__hidden int foo();' in file2.bpf.c
That's safer than traditional linker and the same approach can be
applied to vars.
So externing of __hidden vars won't be possible, but they will name conflict.

2. emit statics when they don't conflict and fail skel gen where there
is a naming conflict.
That helps a bit, but library authors still have to be careful with
both static and global names.
Which is more annoying than traditional C.

3. do #2 style of failing skel gen if there is a naming conflict, but
also introduce namespacing concept, so that both global and static
vars can be automatically namespaced.
That's the proposal above.
This way, I'm guessing, some libraries will use namespaces to avoid
prefixing everything.
The folks that hate namespaces and #pragmas will do manual prefixes for
both static and global vars.

For approaches
char library[]="lru";'
and
#pragma comment(lib, "lru")
the scope of namespace is the whole .bpf.c file.
The clang/llvm already support it, so the job of name mangling would
belong to linker.

For __attribute__((annotate("lib=lru"))) the scope could be any number
of lines in C files between pragma push/pop and can be nested.
This attribute is supported by clang, but not in the bpf backend.
The llvm would prefix both global and static names
in elf file and in btf.
If another file.bpf.c needs to call a function from namespace "lru"
it would need to prefix such a call.
The skel gen job would be #2 above (emit both static and globals if
they don't conflict).
Such namespacing concept would be the closest to c++ namespaces.

If I understood what folks were saying no one is excited about namespaces in C.
So probably #3 is out and sounds like 1 is prefered?
So don't emit statics ?

I'm in favor of not emitting statics. They just seem to cause more
problems than providing any extra benefits. Given it's trivial to just
use globals instead and global vs static is already an explicit signal
of what has to be in BPF skeleton and what's not. See my RFC about
__internal + __hidden semantics, but even if we supported nothing but
STV_DEFAULT globals wouldn't be horrible. Clearly we'd expect users to
just not mess with BPF library's internal state with not way to
enforce that, so I'd still like to have some enforceability.

And keep in mind, BPF code can still use static variables freely, they
will just be BPF-side-only.

Now as for namespacing. Without statics being exposed there doesn't
seem to be any need for namespacing, as globals are supposed to be in
a global namespace with no name mangling and we rely on that guarantee
for BPF skeletons and libbpf APIs that perform lookups by name. But
there is another reason for having a library identifier, though, and
I'll get to it in a second. But before that, between pragma,
annotation and special variable I'd vote for special variable because
  - it's kind of familiar mechanism (like, license and version);
  - it has clear file-wide semantics;
  - as opposed to annotation and pragma, there is a clear mechanism of
preserving such data in ELF, so that linker and/or BPF skeleton
generator can access them.

Of course all of those issues can be overcome, but if we already have
the C/ELF mechanism, I don't see much point in inventing something
new.

Now back to why we might want/need library identifier. Shortly, for
simpler (especially light ) sub-skeleton support.

My original approach for sub-skeleton was to keep BPF library's BTF
information to later (at runtime) find location of .data, .rodata,
.bss data sections (and potentially more). It would work reliably and
would be pretty simple to support through libbpf, because all of the
supporting code is already there in libbpf.

But this is hugely inconvenient for lightweight sub-skeleton, because
it doesn't have libbpf APIs at its disposal.

So my proposal is to allow having a special "library identifier"
variable, e.g., something like:

SEC(".lib") static char[] lib_name = "my_fancy_lib"; (let's not
bikeshed naming for now)

There are complications in how this lib_name is defined across all
files that form the BPF library. It's a bit too obscure to go into
details right now, though. But it feels like need two separate linking
modes: one for linking BPF library and one for a BPF application. But
let's assume for now that we resolve that somehow.

With such library identifiers, BPF static linker will:
  1) enforce uniqueness of library names when linking together
multiple libraries
  2) append zero-size markers to the very beginning and the very end
of each BPF library's DATASECS, something like

DATASEC .data:

   void *___my_fancy_lib__start[0];
   /* here goes .data variables */
   void *___my_fancy_lib__end[0];

And so on for each DATASEC. What those markers provide? Two things:

1). It makes it much easier for sub-skeleton to find where a
particular BPF library's .data/.rodata/.bss starts within the final
BPF application's  .data/.rodata/.bss section. All that without
storing local BTF and doing type comparisons. Only a simple loop over
DATASECs and its variables is needed.

2). (optionally) we can exclude everything between ___<libname>__start
and ___<libname>__end from BPF application's skeleton. It's optional
because I can see cases where BPF application might prefer to have its
own "internal BPF libraries'" state in the final BPF skeleton, instead
of keeping track of many small BPF library sub-skeletons. But for
"external" BPF libraries, it would eliminate accidental corruption of
BPF library state.

I'd actually hold off on implementing #2, because as I said it's not
always clear what's best. But at least that would give us a reliable
and clear mechanism.

It's a pretty long email already and there are a lot things to unpack,
and still more nuances to discuss. So I'll put up BPF static linking +
BPF libraries topic for this week's BPF office hours for anyone
interested to join the live discussion. It would hopefully allow
everyone to get more up to speed and onto the same page on this topic.
But still curious WDYT?

Daniel, Lorenz, John,

what's your take ?