On Fri, Oct 9, 2020 at 7:53 AM Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
>
> Em Thu, Oct 08, 2020 at 04:39:53PM -0700, Andrii Nakryiko escreveu:
> > From: Andrii Nakryiko <andriin@xxxxxx>
> >
> > Switch BTF loading to completely use libbpf's own struct btf and related APIs.
> > BTF encoding is still happening with pahole's own code, so these two code
> > paths are not sharing anything now. String fetching is happening based on
> > whether btfe->strings were set to non-NULL pointer by btf_encoder.
>
> While testing this one I noticed a problem, but it isn't caused by this
> particular patch:
>
> [acme@five pahole]$ cp ~/git/build/v5.9-rc6+/net/ipv4/tcp_ipv4.o .
> [acme@five pahole]$ readelf -SW tcp_ipv4.o | grep BTF
> [acme@five pahole]$ pahole -J tcp_ipv4.o
> [acme@five pahole]$ readelf -SW tcp_ipv4.o | grep BTF
> [105] .BTF PROGBITS 0000000000000000 0fbb3c 03f697 00 0 0 1
> [acme@five pahole]$ ./btfdiff tcp_ipv4.o
> --- /tmp/btfdiff.dwarf.BDAvGi 2020-10-09 11:41:45.161509391 -0300
> +++ /tmp/btfdiff.btf.p81icw 2020-10-09 11:41:45.177509720 -0300
> @@ -4056,7 +4056,7 @@ struct tcp_congestion_ops {
> u32 (*min_tso_segs)(struct sock *); /* 96 8 */
> u32 (*sndbuf_expand)(struct sock *); /* 104 8 */
> void (*cong_control)(struct sock *, const struct rate_sample *); /* 112 8 */
> - size_t (*get_info)(struct sock *, u32, int *, union tcp_cc_info *); /* 120 8 */
> + size_t (*get_info)(struct sock *, u32, int *, struct tcp_cc_info *); /* 120 8 */
It's a bug in btf_loader.c. When loading BTF_KIND_FWD, we always
assume it's struct forward reference. But I checked, generated BTF
does have the valid info (kind_flag), so it should be easy to fix,
I'll send a fix.
The reason we don't see it in vmlinux is because after dedup that
forward reference is resolved to a full type, which is clearly union.
[...]
