Re: Build failures: unresolved symbol vfs_getattr

[Jiri Olsa <jolsa@redhat.com>]

On Mon, Sep 14, 2020 at 10:48:36AM -0400, Veronika Kabatova wrote:
> 
> Hello,
> 
> we tested the bpf-next tree with CKI and ran across build failures. The
> important part of the build log is:
> 
> 00:18:05   GEN     .version
> 00:18:05   CHK     include/generated/compile.h
> 00:18:05   LD      vmlinux.o
> 00:18:27   MODPOST vmlinux.symvers
> 00:18:27   MODINFO modules.builtin.modinfo
> 00:18:27   GEN     modules.builtin
> 00:18:27   LD      .tmp_vmlinux.btf
> 00:18:42   BTF     .btf.vmlinux.bin.o
> 00:19:13   LD      .tmp_vmlinux.kallsyms1
> 00:19:19   KSYM    .tmp_vmlinux.kallsyms1.o
> 00:19:22   LD      .tmp_vmlinux.kallsyms2
> 00:19:25   KSYM    .tmp_vmlinux.kallsyms2.o
> 00:19:28   LD      vmlinux
> 00:19:40   BTFIDS  vmlinux
> 00:19:40 FAILED unresolved symbol vfs_getattr
> 00:19:40 make[2]: *** [Makefile:1167: vmlinux] Error 255
> 00:19:40 make[1]: *** [scripts/Makefile.package:109: targz-pkg] Error 2
> 00:19:40 make: *** [Makefile:1528: targz-pkg] Error 2

hi,
it looks like broken BTF data to me, I checked that build
and found we have multiple records for functions, like
for filp_close:

	[23381] FUNC_PROTO '(anon)' ret_type_id=19 vlen=2
		'(anon)' type_id=464
		'id' type_id=960
	[23382] FUNC 'filp_close' type_id=23381 linkage=static


	[33073] FUNC_PROTO '(anon)' ret_type_id=19 vlen=2
		'filp' type_id=464
		'id' type_id=960
	[33074] FUNC 'filp_close' type_id=33073 linkage=static


or vfs_getattr:

	[33513] FUNC_PROTO '(anon)' ret_type_id=19 vlen=4
		'path' type_id=741
		'stat' type_id=1095
		'request_mask' type_id=29
		'query_flags' type_id=8

	[33514] FUNC 'vfs_getattr' type_id=33513 linkage=static

	[1094] FUNC_PROTO '(anon)' ret_type_id=19 vlen=4
		'(anon)' type_id=741
		'(anon)' type_id=1095
		'(anon)' type_id=29
		'(anon)' type_id=8

	[35099] FUNC 'vfs_getattr' type_id=1094 linkage=static


and because we go through all BTF data until we resolve all we have,
the doubled funcs will screw our internal counter and we skip a function

the change below will workaround that, but I think we should fail in
this case.. if I'm not missing something 2 FUNC records for one function
in BTF data

$ pahole --version
v1.17

HEAD is 2bab48c5b Merge branch 'improve-bpf-tcp-cc-init'

thoughts? thanks
jirka


---
diff --git a/tools/bpf/resolve_btfids/main.c b/tools/bpf/resolve_btfids/main.c
index dfa540d8a02d..a33e56553e52 100644
--- a/tools/bpf/resolve_btfids/main.c
+++ b/tools/bpf/resolve_btfids/main.c
@@ -525,7 +525,7 @@ static int symbols_resolve(struct object *obj)
 		}
 
 		id = btf_id__find(root, str);
-		if (id) {
+		if (id && !id->id) {
 			id->id = type_id;
 			(*nr)--;
 		}