Re: [PATCH bpf 1/2] libbpf: Fix stale raw data after btf__relocate()

From: Alexei Starovoitov

Date: Sun Nov 01 2026 - 13:03:12 EST


On Sun, Nov 01, 2026 at 07:56 AM Jiayuan Chen <jiayuan.chen@xxxxxxxxx> wrote:
> Call btf_ensure_modifiable() first, like other BTF mutators do, so the
> cache is invalidated.

btf_parse_elf() calls btf__relocate() for every module with .BTF.base.
That BTF is read-only. types_data points into raw_data, so raw_data is
not a cache there. Only raw_data_swapped can be stale.
With this patch that path changes:

today with the patch
types rewritten in place copied
strings as-is copied and hashed into strset
named_start_id kept 0, lookup by name is linear
has_hdr_extra relocates -EOPNOTSUPP

The last row: btf__parse_split() of a module with new header fields fails.
Ihor removes the same side effect from btf__find_str():
https://lore.kernel.org/bpf/20261030233355.1836749-2-ihor.solodrai@xxxxxxxxx/

can we drop only the copies ?
The kernel calls btf_check_sorted() after btf_relocate() too.

Something like this (not even compile tested):

diff --git a/tools/lib/bpf/btf.c b/tools/lib/bpf/btf.c
index d2ff6f5..d67f78d 100644
--- a/tools/lib/bpf/btf.c
+++ b/tools/lib/bpf/btf.c
@@ -2033,7 +2033,8 @@ struct btf *btf__load_from_kernel_by_id(__u32 id)

static void btf_invalidate_raw_data(struct btf *btf)
{
- if (btf->raw_data)
+ /* types_data and strs_data of read-only BTF point into raw_data */
+ if (btf->raw_data && btf_is_modifiable(btf))
btf_free_raw_data(btf);
if (btf->raw_data_swapped) {
free(btf->raw_data_swapped);
@@ -2099,11 +2100,11 @@ static int btf_ensure_modifiable(struct btf *btf)
if (!btf->base_btf && btf->hdr.str_len == 1)
btf->strs_deduped = true;

+ btf->modifiable = true;
+
/* invalidate raw_data representation */
btf_invalidate_raw_data(btf);

- btf->modifiable = true;
-
return 0;

err_out:
@@ -6642,8 +6643,12 @@ int btf__relocate(struct btf *btf, const struct btf *base_btf)
{
int err = btf_relocate(btf, base_btf, NULL, NULL);

- if (!err)
+ /* types are rewritten in place, some of them even on error */
+ btf_invalidate_raw_data(btf);
+ if (!err) {
btf->owns_base = false;
+ btf_check_sorted(btf);
+ }
return libbpf_err(err);
}


Also pls target bpf-next. The selftest is contrived.

pw-bot: cr