On Thu, Jun 24, 2021 at 8:06 AM Daniel Borkmann [off-list ref] wrote:

On 6/24/21 6:03 AM, Shuyi Cheng wrote:

quoted

In order to enable the older kernel to use the CO-RE feature, load the
vmlinux btf of the specified path.

Learn from Andrii's comments in [0], add the custom_btf_path parameter
to bpf_obj_open_opts, you can directly use the skeleton's
<objname>_bpf__open_opts function to pass in the custom_btf_path
parameter.

Prior to this, there was also a developer who provided a patch with
similar functions. It is a pity that the follow-up did not continue to
advance. See [1].

      [0]https://lore.kernel.org/bpf/CAEf4BzbJZLjNoiK8_VfeVg_Vrg=9iYFv+po-38SMe=UzwDKJ=Q@mail.gmail.com/#t (local)
      [1]https://yhbt.net/lore/all/CAEf4Bzbgw49w2PtowsrzKQNcxD4fZRE6AKByX-5-dMo-+oWHHA@mail.gmail.com/

Signed-off-by: Shuyi Cheng <redacted>
---
  tools/lib/bpf/libbpf.c | 23 ++++++++++++++++++++---
  tools/lib/bpf/libbpf.h |  6 +++++-
  2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
index 1e04ce7..518b19f 100644
--- a/tools/lib/bpf/libbpf.c
+++ b/tools/lib/bpf/libbpf.c

@@ -509,6 +509,8 @@ struct bpf_object {
      void *priv;
      bpf_object_clear_priv_t clear_priv;

+     char *custom_btf_path;
+

nit: This should rather go to the 'Parse and load BTF vmlinux if any of [...]'
section of struct bpf_object, and for consistency, I'd keep the btf_ prefix,
like: char *btf_custom_path

quoted

      char path[];
  };
  #define obj_elf_valid(o)    ((o)->efile.elf)

@@ -2679,8 +2681,15 @@ static int bpf_object__load_vmlinux_btf(struct bpf_object *obj, bool force)
      if (!force && !obj_needs_vmlinux_btf(obj))
              return 0;

-     obj->btf_vmlinux = libbpf_find_kernel_btf();
-     err = libbpf_get_error(obj->btf_vmlinux);
+     if (obj->custom_btf_path) {
+             obj->btf_vmlinux = btf__parse(obj->custom_btf_path, NULL);
+             err = libbpf_get_error(obj->btf_vmlinux);
+             pr_debug("loading custom vmlinux BTF '%s': %d\n", obj->custom_btf_path, err);
+     } else {
+             obj->btf_vmlinux = libbpf_find_kernel_btf();
+             err = libbpf_get_error(obj->btf_vmlinux);
+     }

Couldn't we do something like (only compile-tested):

I wonder what are the benefits of this approach, though. My
expectation is that if the user specifies a custom BTF path and BTF is
missing then the whole bpf_object load process should fail, but in
this case it will be silently ignored. Also, if custom BTF is
specified, that custom BTF has to be used even if
/sys/kernel/btf/vmlinux is present, but the patch below will still
prefer /sys/kernel/btf/vmlinux.

So the semantics is different. I'm not saying it's wrong, but I think
it means we need to discuss what behavior we are after first.

[...]