Re: [PATCH bpf-next 0/9] Add support for type-based and enum value-based CO-RE relocations

On Tue, Aug 18, 2020 at 6:37 PM Alexei Starovoitov
[off-list ref] wrote:

On Tue, Aug 18, 2020 at 06:31:51PM -0700, Andrii Nakryiko wrote:

quoted

On Tue, Aug 18, 2020 at 6:21 PM Alexei Starovoitov
[off-list ref] wrote:

quoted

On Tue, Aug 18, 2020 at 03:39:12PM -0700, Andrii Nakryiko wrote:

quoted

This patch set adds libbpf support to two new classes of CO-RE relocations:
type-based (TYPE_EXISTS/TYPE_SIZE/TYPE_ID_LOCAL/TYPE_ID_TARGET) and enum
value-vased (ENUMVAL_EXISTS/ENUMVAL_VALUE):

LLVM patches adding these relocation in Clang:
  - __builtin_btf_type_id() ([0], [1], [2]);
  - __builtin_preserve_type_info(), __builtin_preserve_enum_value() ([3], [4]).

I've applied patches 1-4, since they're somewhat indepedent of new features in 5+.
What should be the process to land the rest?
Land llvm first and add to bpf/README.rst that certain llvm commmits are necessary
to build the tests?

Clang patches landed about two weeks ago, so they are already in Clang
nightly builds. libbpf CI should work fine as it uses clang-12 nightly
builds.

quoted

But CI will start failing. We can wait for that to be fixed,
but I wonder is there way to detect new clang __builtins automatically in
selftests and skip them if clang is too old?

There is a way to detect built-ins availability (__has_builtin macro,
[0]) from C code. If we want to do it from Makefile, though, we'd need
to do feature detection similar to how we did reallocarray and
libbpf-elf-mmap detection I just removed in the other patch set :).
Then we'll also need to somehow blacklist tests. Maintaining that
would be a pain, honestly. So far selftests/bpf assumed the latest
Clang, though, so do you think we should change that, or you were
worried that patches hadn't landed yet?

I was hoping that libbpf.h can have builtins unconditionally, but selftests can
do feature detection automatically and mark them as 'skip'.
People have been forever complaining about constant need to upgrade clang.
In this case I think the feature is not fundamental enough (unlike the first
set of builtins) to force adoption of new clang.
If/when we start using these new builtins beyond selftests
that would be a different story.

Ok, took some tinkering to do this for btf_type_id() tests, but
everything works now as you described above. Posted v2.