Re: [RFC PATCH bpf-next v3 1/2] libbpf: support selective kernel module BTF loading via bpf_object_open_opts

[Andrii Nakryiko]

On Thu, Aug 20, 2026 at 8:51 PM Fuyu Zhao <zhaofuyu@xxxxxxxx> wrote:
>
> Hi Eduard,
>
> Thanks for the suggestions. Please see my responses inline below.
>
> On 8/21/2026 7:58 AM, Eduard Zingerman wrote:
> > On Wed, 2026-08-19 at 17:04 +0800, Fuyu Zhao wrote:
> >
> > Overall the logic seem to be fine for me, please find a few comments below.
> >
> > ...
> >
> >> diff --git a/tools/lib/bpf/libbpf.c b/tools/lib/bpf/libbpf.c
> >> index 514e4e9daa82..37934ca49dd7 100644
> >> --- a/tools/lib/bpf/libbpf.c
> >> +++ b/tools/lib/bpf/libbpf.c
> >> @@ -779,6 +779,13 @@ struct bpf_object {
> >>      char *token_path;
> >>      int token_fd;
> >>
> >> +    /* kernel module BTFs to load, as specified via bpf_object_open_opts */
> >> +    struct {
> >> +            char **names;
> >> +            size_t nr_names;
> >> +            struct hashmap *names_map;
> >
> > We already have a strset type, please use it instead of a direct
> > hashmap reference. This would remove the need for `names` field
> > and simplify the bpf_object__init_kmod_btfs() function.
> >
> > Also, do you expect `nr_names` to be high?
> > If not, wouldn't plain array search be simpler/faster here?
> >
>
> I initially used a hashmap to avoid repeatedly walking the requested
> module name list. But I don't expect nr_names to be high in practice, so
> a plain array search should be sufficient and simpler. I'll change this
> in the next version.
>
> >> +    } *kmod_btfs;
> >
> > Why indirection?
> >
>
> I followed the existing pattern in bpf_object and used an anonymous
> structure with indirection here. I don't have a strong reason to keep
> the indirection, though. I'll simplify it.
>
> > Also, I agree with the bot here, prior fields use "module" in the name,
> > so something like "btf_module_names" or similar is a better fit.
> > Same for publicly visible 'opts' name.
> >
>
> I think `btf_module_names` is a much better name than my original choice.
> I'll change that in `opts`, too.
>
> > ...
> >
> >> +static int bpf_object__init_kmod_btfs(struct bpf_object *obj,
> >> +                                  const struct bpf_object_open_opts *opts)
> >> +{
> >> +    const char **kmod_btf_names;
> >> +    size_t i, kmod_btf_names_cnt;
> >> +    int err;
> >> +
> >> +    kmod_btf_names = OPTS_GET(opts, kmod_btf_names, NULL);
> >> +    if (!kmod_btf_names)
> >> +            return 0;
> >> +
> >> +    kmod_btf_names_cnt = OPTS_GET(opts, kmod_btf_names_cnt, 0);
> >> +    if (!kmod_btf_names_cnt) {
> >> +            pr_warn("kmod_btf_names_cnt must be set when kmod_btf_names is provided\n");
> >> +            return -EINVAL;
> >
> > I kinda agree with the bot here, why disallow an empty filter here?
> >
>
> I understand the point now. I hadn't considered the use case of
> explicitly skipping all module BTFs. A non-NULL `kmod_btf_names` with a
> zero count is a way to represent this case, so I'll support it in the
> next version.
>

besides all Eduard's suggestions, one naming nit: don't use double
underscore naming, those are reserved for extern method-like APIs,
while you are adding internal helpers


> >> +    }
> >> +
> >> +    obj->kmod_btfs = calloc(1, sizeof(*obj->kmod_btfs));
> >> +    if (!obj->kmod_btfs)
> >> +            return -ENOMEM;
> >> +
> >> +    obj->kmod_btfs->names = calloc(kmod_btf_names_cnt, sizeof(char *));
> >> +    if (!obj->kmod_btfs->names) {
> >> +            err = -ENOMEM;
> >> +            goto err_out;
> >> +    }
> >> +
> >> +    obj->kmod_btfs->names_map = hashmap__new(mod_name_hash_fn,
> >> +                                             mod_name_equal_fn, NULL);
> >> +    if (IS_ERR(obj->kmod_btfs->names_map)) {
> >> +            err = PTR_ERR(obj->kmod_btfs->names_map);
> >> +            obj->kmod_btfs->names_map = NULL;
> >> +            goto err_out;
> >> +    }
> >> +
> >> +    for (i = 0; i < kmod_btf_names_cnt; i++) {
> >> +            size_t idx = obj->kmod_btfs->nr_names;
> >> +
> >> +            if (!kmod_btf_names[i] || !kmod_btf_names[i][0]) {
> >> +                    pr_warn("invalid kernel module BTF name at index %zu\n", i);
> >> +                    err = -EINVAL;
> >> +                    goto err_out;
> >> +            }
> >> +
> >> +            obj->kmod_btfs->names[idx] = strdup(kmod_btf_names[i]);
> >> +            if (!obj->kmod_btfs->names[idx]) {
> >> +                    err = -ENOMEM;
> >> +                    goto err_out;
> >> +            }
> >> +
> >> +            err = hashmap__add(obj->kmod_btfs->names_map,
> >> +                               obj->kmod_btfs->names[idx], 0);
> >> +            if (err) {
> >> +                    zfree(&obj->kmod_btfs->names[idx]);
> >> +                    if (err == -EEXIST) {
> >> +                            pr_warn("duplicate kmod BTF name '%s' ignored\n",
> >> +                                    kmod_btf_names[i]);
> >
> > Nit: I'd downgrade this to debug level, if at all.
> >
>
> Agreed. I'll downgrade this to debug level in the next version.
>

when you copy over a list of kernel module names, sort it, and do a
pass to detect duplicates. if there are any -- just error out, specify
in the documentation that the list of kernel modules names should not
contain duplicates. It's not a particularly limiting restriction, is
it?

> >> +                            continue;
> >> +                    }
> >> +                    goto err_out;
> >> +            }
> >> +            obj->kmod_btfs->nr_names++;
> >> +    }
> >> +    return 0;
> >> +
> >> +err_out:
> >> +    bpf_object__free_kmod_btfs(obj);
> >> +    return err;
> >> +}
> >
> > ...
> >
> >> @@ -8515,6 +8645,7 @@ static struct bpf_object *bpf_object_open(const char *path, const void *obj_buf,
> >>      err = err ? : bpf_object__init_maps(obj, opts);
> >>      err = err ? : bpf_object_init_progs(obj, opts);
> >>      err = err ? : bpf_object__collect_relos(obj);
> >> +    err = err ? : bpf_object__init_kmod_btfs(obj, opts);
> >
> > Nit: all other options are collected before bpf_object__elf_init()
> >      call just above.
> >
>
> Agreed. I'll change it in the next version.
>
> >>      if (err)
> >>              goto out;
> >>
> >
> > ...
> >
> >> diff --git a/tools/lib/bpf/libbpf.h b/tools/lib/bpf/libbpf.h
> >> index b965ad571540..2f8ff6d2d3df 100644
> >> --- a/tools/lib/bpf/libbpf.h
> >> +++ b/tools/lib/bpf/libbpf.h
> >> @@ -224,10 +224,24 @@ struct bpf_object_open_opts {
> >>       * point (/sys/fs/bpf), in case this default behavior is undesirable.
> >>       */
> >>      const char *bpf_token_path;
> >> +    /*
> >> +     * Optional list of kernel module names whose BTFs should be loaded.
> >> +     * kmod_btf_names_cnt specifies the number of entries in
> >> +     * kmod_btf_names.
> >> +     *
> >> +     * If kmod_btf_names is NULL, all module BTFs are loaded, preserving
> >> +     * the default behavior. Otherwise, only the specified module BTFs
> >> +     * are loaded.
> >> +     *
> >> +     * kmod_btf_names_cnt must be non-zero when kmod_btf_names is
> >> +     * non-NULL; otherwise -EINVAL is returned.
> >
> > Nit: please add a complete list of behaviors this affects, e.g. CO-RE,
> >      fentry/fexit resolution, etc.
> >
>
> I'll expand the comment to cover the affected behaviors in the next version.
>
> >> +     */
> >> +    const char **kmod_btf_names;
> >> +    size_t kmod_btf_names_cnt;
> >>
> >>      size_t :0;
> >>  };
> >> -#define bpf_object_open_opts__last_field bpf_token_path
> >> +#define bpf_object_open_opts__last_field kmod_btf_names_cnt
> >>
> >>  /**
> >>   * @brief **bpf_object__open()** creates a bpf_object by opening
>