On Thu, Aug 29, 2024 at 06:40:59PM -0700, Eduard Zingerman wrote:
> On Thu, 2024-08-29 at 18:27 -0700, Tony Ambardar wrote:
>
>
> > Thanks for looking at this. I ran into the CI failure while using s390x
> > to test a series adding libbpf bi-endian support. Since I'm deep into
> > endianness issues right now, I thought to try the fix you suggested just
> > to make some progress but noticed the CI failure has disappeared.[0]
>
> Hi Tony,
>
> There is no fix yet, sorry :)
> I think that something like below should do the trick:
>
> --- a/tools/lib/bpf/btf.c
> +++ b/tools/lib/bpf/btf.c
> @@ -5394,6 +5394,7 @@ int btf__distill_base(const struct btf *src_btf, struct btf **new_base_btf,
> new_base = btf__new_empty();
> if (!new_base)
> return libbpf_err(-ENOMEM);
> + btf__set_endianness(new_base, btf__endianness(src_btf));
> dist.id_map = calloc(n, sizeof(*dist.id_map));
> if (!dist.id_map) {
> err = -ENOMEM;
>
> as far as I understand btf__raw_data() should do all conversions after this.
> But I have not tested it yet and would be AFK for a few hours.
>
Hi Eduard,
Yes, btf__raw_data() will work as expected.
I updated my local pahole and managed to reproduce the problem after
cross-compiling to s390x. Looking at lib/bpf/btf.c, I see one more spot
that needs to preserve source endianness compared to patch above, and
local testing under QEMU now works for me:
root@(none):/usr/libexec/kselftests-bpf# insmod bpf_testmod.ko
bpf_testmod: loading out-of-tree module taints kernel.
bpf_testmod: module verification failed: signature and/or required key
missing - tainting kernel
root@(none):/usr/libexec/kselftests-bpf# ./test_progs -a map_ptr
#166 map_ptr:OK
Summary: 1/0 PASSED, 0 SKIPPED, 0 FAILED
and:
$ llvm-readelf -x .BTF
bpf_testmod.ko | head -5
Hex dump of section '.BTF':
0x00000000 eb9f0100 00000018 00000000 00001a30 ...............0
0x00000010 00001a30 00001180 00000000 0a000000 ...0............
0x00000020 00000022 00000000 03000000 00000000 ..."............
0x00000030 00000028 00000006 0000002b 00000000 ...(.......+....
$ llvm-readelf -x .BTF.base
bpf_testmod.ko | head -5
Hex dump of section '.BTF.base':
0x00000000 eb9f0100 00000018 00000000 000001fc ................
0x00000010 000001fc 000001ea 00000001 01000000 ................
0x00000020 00000008 00000040 00000013 01000000 .......@........
0x00000030 00000001 00000008 00000018 01000000 ................
Please try with the patch below, or I can just send a proper one to the
list with some added "Co-developed-by:" if easier?
--- a/tools/lib/bpf/btf.c
+++ b/tools/lib/bpf/btf.c
@@ -996,6 +996,7 @@ static struct btf *btf_new_empty(struct btf *base_btf)
btf->base_btf = base_btf;
btf->start_id = btf__type_cnt(base_btf);
btf->start_str_off = base_btf->hdr->str_len;
+ btf->swapped_endian = base_btf->swapped_endian;
}
/* +1 for empty string at offset 0 */
@@ -5554,6 +5555,7 @@ int btf__distill_base(const struct btf *src_btf,
struct btf **new_base_btf,
new_base = btf__new_empty();
if (!new_base)
return libbpf_err(-ENOMEM);
+ btf__set_endianness(new_base, btf__endianness(src_btf));
dist.id_map = calloc(n, sizeof(*dist.id_map));
if (!dist.id_map) {
err = -ENOMEM;
> > Did something get fixed already? I can't seem to find the change.
>
> pahole version w/o support for distilled base was pinned on CI:
> https://github.com/kernel-patches/vmtest/pull/285/commits/d3eff26fc978ca8fb3bce3f93421f7425aef0f55
>
Ah, got it! That makes more sense now. Thanks for the extra details.
Take care,
Tony
>
> Thanks,
> Eduard
