On Tue, 23 Apr 2024 13:12:04 -0700 Yonghong Song <yonghong.song@xxxxxxxxx> wrote:
> On 4/23/24 6:15 AM, Xin Liu wrote:
> > On Mon, 22 Apr 2024 10:43:38 -0700 Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx> wrote:
> >
> >> On Mon, Apr 22, 2024 at 7:46 AM Xin Liu <liuxin350@xxxxxxxxxx> wrote:
> >>> In btf__add_int, the size of the new btf_kind_int type is limited.
> >>> When the size is greater than 16, btf__add_int fails to be added
> >>> and -EINVAL is returned. This is usually effective.
> >>>
> >>> However, when the built-in type __builtin_aarch64_simd_xi in the
> >>> NEON instruction is used in the code in the arm64 system, the value
> >>> of DW_AT_byte_size is 64. This causes btf__add_int to fail to
> >>> properly add btf information to it.
> >>>
> >>> like this:
> >>> ...
> >>> <1><cf>: Abbrev Number: 2 (DW_TAG_base_type)
> >>> <d0> DW_AT_byte_size : 64 // over max size 16
> >>> <d1> DW_AT_encoding : 5 (signed)
> >>> <d2> DW_AT_name : (indirect string, offset: 0x53): __builtin_aarch64_simd_xi
> >>> <1><d6>: Abbrev Number: 0
> >>> ...
> >>>
> >>> An easier way to solve this problem is to treat it as a base type
> >>> and set byte_size to 64. This patch is modified along these lines.
> >>>
> >>> Fixes: 4a3b33f8579a ("libbpf: Add BTF writing APIs")
> >>> Signed-off-by: Xin Liu <liuxin350@xxxxxxxxxx>
> >>> ---
> >>> tools/lib/bpf/btf.c | 2 +-
> >>> 1 file changed, 1 insertion(+), 1 deletion(-)
> >>>
> >>> diff --git a/tools/lib/bpf/btf.c b/tools/lib/bpf/btf.c
> >>> index 2d0840ef599a..0af121293b65 100644
> >>> --- a/tools/lib/bpf/btf.c
> >>> +++ b/tools/lib/bpf/btf.c
> >>> @@ -1934,7 +1934,7 @@ int btf__add_int(struct btf *btf, const char *name, size_t byte_sz, int encoding
> >>> if (!name || !name[0])
> >>> return libbpf_err(-EINVAL);
> >>> /* byte_sz must be power of 2 */
> >>> - if (!byte_sz || (byte_sz & (byte_sz - 1)) || byte_sz > 16)
> >>> + if (!byte_sz || (byte_sz & (byte_sz - 1)) || byte_sz > 64)
> >>
> >> maybe we should just remove byte_sz upper limit? We can probably
> >> imagine 256-byte integers at some point, so why bother artificially
> >> restricting it?
> >>
> >> pw-bot: cr
> > In the current definition of btf_kind_int, bits has only 8 bits, followed
> > by 8 bits of unused interval. When we expand, we should only use 16 bits
> > at most, so the maximum value should be 8192(1 << 16 / 8), directly removing
> > the limit of byte_sz. It may not fit the current design. For INT type btfs
> > greater than 255, how to dump is still a challenge.
>
> Looking at this patch. Now I remember that I have an old pahole patch
> to address similar issues
> https://lore.kernel.org/bpf/20230426055030.3743074-1-yhs@xxxxxx/
> which is not merged and I forgot that.
>
> In that particular case, the int size is 1024 bytes.
> Currently the int type more than 16 bytes cannot be dumped in libbpf.
> Do you have a particular use case to use your__builtin_aarch64_simd_xi() type
> in bpf program? I guess probably not as BPF does not support
> builtin function your__builtin_aarch64_simd_xi().
>
Currently, there is no use case of byte_sz in btf, so let's remove
__builtin_aarch64_simd_xi first.At least this will support the kernel
compilation phase without causing the kernel to fail directly when
generating btf.
> >
> > Does the current version support a maximum of 8192 bytes?
> >
> >>> return libbpf_err(-EINVAL);
> >>> if (encoding & ~(BTF_INT_SIGNED | BTF_INT_CHAR | BTF_INT_BOOL))
> >>> return libbpf_err(-EINVAL);
> >>> --
> >>> 2.33.0
> >>>
