On Thu, Dec 19, 2024 at 02:43:30PM -0800, Martin KaFai Lau wrote:
> On 12/17/24 6:44 PM, D. Wythe wrote:
> >Here are four possible case:
> >
> >+--------+-------------+-------------+---------------------------------+
> >| | st_opx_xxx | xxx | |
> >+--------+-------------+-------------+---------------------------------+
> >| case 0 | btf_vmlinux | bft_vmlinux | be used and reg only in vmlinux |
> >+--------+-------------+-------------+---------------------------------+
> >| case 1 | btf_vmlinux | bpf_mod | INVALID |
> >+--------+-------------+-------------+---------------------------------+
> >| case 2 | btf_mod | btf_vmlinux | reg in mod but be used both in |
> >| | | | vmlinux and mod. |
> >+--------+-------------+-------------+---------------------------------+
> >| case 3 | btf_mod | btf_mod | be used and reg only in mod |
> >+--------+-------------+-------------+---------------------------------+
> >
> >At present, cases 0, 1, and 3 can be correctly identified, because
> >st_ops_xxx is searched from the same btf with xxx. In order to
> >handle case 2 correctly without affecting other cases, we cannot simply
> >change the search method for st_ops_xxx from find_btf_by_prefix_kind()
> >to find_ksym_btf_id(), because in this way, case 1 will not be
> >recognized anymore.
> > snprintf(tname, sizeof(tname), "%.*s",
> >@@ -1020,17 +1021,25 @@ find_struct_ops_kern_types(struct bpf_object *obj, const char *tname_raw,
> > }
> > kern_type = btf__type_by_id(btf, kern_type_id);
> >+ ret = snprintf(stname, sizeof(stname), "%s%s", STRUCT_OPS_VALUE_PREFIX, tname);
>
> How about always look for "struct bpf_struct_ops_smc_ops" first,
> figure out the btf, and then look for "struct smc_ops", would it
> work?
I think this might not work, as the core issue lies in the fact that
bpf_struct_ops_smc_ops and smc_ops are located on different btf.
Searching for one fisrt cannot lead to the inference of the other.
>
> If CONFIG_SMC=y instead of =m, this change cannot be tested?
>
That is indeed a problem, but currently there is no better solution
unless the CI can add a step to run 'make modules_install'.
Best wishes,
D. Wythe
> >+ if (ret < 0 || ret >= sizeof(stname))
> >+ return -ENAMETOOLONG;
> >+
> > /* Find the corresponding "map_value" type that will be used
> > * in map_update(BPF_MAP_TYPE_STRUCT_OPS). For example,
> > * find "struct bpf_struct_ops_tcp_congestion_ops" from the
> > * btf_vmlinux.
> > */
> >- kern_vtype_id = find_btf_by_prefix_kind(btf, STRUCT_OPS_VALUE_PREFIX,
> >- tname, BTF_KIND_STRUCT);
> >+ kern_vtype_id = btf__find_by_name_kind(btf, stname, BTF_KIND_STRUCT);
> > if (kern_vtype_id < 0) {
> >- pr_warn("struct_ops init_kern: struct %s%s is not found in kernel BTF\n",
> >- STRUCT_OPS_VALUE_PREFIX, tname);
> >- return kern_vtype_id;
> >+ if (kern_vtype_id == -ENOENT && !*mod_btf)
> >+ kern_vtype_id = find_ksym_btf_id(obj, stname, BTF_KIND_STRUCT,
> >+ &btf, mod_btf);
> >+ if (kern_vtype_id < 0) {
> >+ pr_warn("struct_ops init_kern: struct %s is not found in kernel BTF\n",
> >+ stname);
> >+ return kern_vtype_id;
> >+ }
> > }
> > kern_vtype = btf__type_by_id(btf, kern_vtype_id);
> >@@ -1046,8 +1055,8 @@ find_struct_ops_kern_types(struct bpf_object *obj, const char *tname_raw,
> > break;
> > }
> > if (i == btf_vlen(kern_vtype)) {
> >- pr_warn("struct_ops init_kern: struct %s data is not found in struct %s%s\n",
> >- tname, STRUCT_OPS_VALUE_PREFIX, tname);
> >+ pr_warn("struct_ops init_kern: struct %s data is not found in struct %s\n",
> >+ tname, stname);
> > return -EINVAL;
> > }
