Re: [PATCH bpf-next v1 06/14] resolve_btfids: Discover kfuncs from BTF ID sets

On 6/3/26 4:45 PM, Andrii Nakryiko wrote:
> On Mon, Jun 1, 2026 at 3:19 PM Ihor Solodrai <[email protected]> wrote:
>>
>> collect_kfuncs() currently uses bpf_kfunc decl tags to identify the
>> list of kfuncs. The decl tags are generated by pahole, which makes
>> current implementation implicitly rely on those tags being generated.
>>
>> The authoritative source, used by the the BPF verifier for kfunc
>> registration, of functions being BPF kfuncs are
>> BTF_KFUNCS_START()/END() declarations. These are BTF_ID_SET8 under the
>> hood. Currently resolve_btfids reads kfunc flags from these sets, and
>> populates them with BTF IDs.
>>
>> Implement kfunc discovery from BTF_ID_SET8 symbols in resolve_btfids,
>> removing the dependency on pahole's emmission of decl tags.
>>
>> Walk BTF_ID_KIND_SET8 sets, and use the address-to-symbol index to
>> look up set entry's BTF_ID symbol name (before .BTF_ids is patched),
>> recording the paired flags directly. This makes find_kfunc_flags()
>> helper unnecessary, so it's removed.
>>
>> Kernel functions can appear in more than one set, which is legitimate,
>> since kfunc sets are prog-type dependent in the kernel. So for btf2btf
>> processing deduplicate kfuncs by BTF ID, accumulate (OR) the flags,
>> and warn on flags mismatch to catch inconsistent declarations.
>>
>> Signed-off-by: Ihor Solodrai <[email protected]>
>> ---
>>  tools/bpf/resolve_btfids/main.c | 122 ++++++++++++++------------------
>>  1 file changed, 55 insertions(+), 67 deletions(-)
>>
>> diff --git a/tools/bpf/resolve_btfids/main.c b/tools/bpf/resolve_btfids/main.c
>> index 43512af13148..d35a7b2460e8 100644
>> --- a/tools/bpf/resolve_btfids/main.c
>> +++ b/tools/bpf/resolve_btfids/main.c
>> @@ -970,6 +970,23 @@ static int push_kfunc(struct btf2btf_context *ctx, struct kfunc *kfunc)
>>         struct kfunc *arr = ctx->kfuncs;
>>         u32 cap = ctx->max_kfuncs;
>>
>> +       /*
>> +        * A kfunc can be listed in multiple BTF ID sets.
>> +        * In this case, dedup by btf_id and accumulate kfunc flags.
>> +        */
>> +       for (u32 i = 0; i < ctx->nr_kfuncs; i++) {
>> +               if (ctx->kfuncs[i].btf_id != kfunc->btf_id)
>> +                       continue;
>> +
> 
> with hundreds of kfuncs, this O(N^2) approach is going to be a bit
> slow, should we use rb tree for lookups?

Yeah, it's O(n^2). I was thinking it's fine, because the number of
kfuncs is ~ a few hundreds. I'll use the rbtree in v2.

> 
>> +               if (ctx->kfuncs[i].flags != kfunc->flags) {
>> +                       pr_err("WARN: resolve_btfids: inconsistent flags for kfunc %s: 0x%x != 0x%x\n",
>> +                              kfunc->name, ctx->kfuncs[i].flags, kfunc->flags);
>> +                       warnings++;
>> +               }
>> +               ctx->kfuncs[i].flags |= kfunc->flags;
>> +               return 0;
>> +       }
>> +
>>         if (ctx->nr_kfuncs + 1 > cap) {
>>                 cap = max(cap + 256, cap * 2);
>>                 arr = realloc(arr, sizeof(struct kfunc) * cap);
> 
> [...]