On 10/7/22 1:21 PM, Arnaldo Carvalho de Melo wrote:
Em Thu, Oct 06, 2022 at 02:23:16PM -0300, Arnaldo Carvalho de Melo escreveu:
Em Thu, Oct 06, 2022 at 09:04:58AM -0700, Andrii Nakryiko escreveu:
On Thu, Oct 6, 2022 at 7:00 AM Arnaldo Carvalho de Melo <arnaldo.melo@xxxxxxxxx> wrote:
Em Thu, Oct 06, 2022 at 10:43:22AM -0300, Arnaldo Carvalho de Melo escreveu:
15e20ce2324a:~/git/linux # readelf -wi ./arch/x86/entry/entry.o
Contents of the .debug_info section:
Compilation Unit @ offset 0:
Length: 0x35 (32-bit)
Version: 5
Unit Type: DW_UT_compile (1)
Abbrev Offset: 0
Pointer Size: 8
<0><c>: Abbrev Number: 1 (DW_TAG_compile_unit)
<d> DW_AT_stmt_list : 0
<11> DW_AT_low_pc : 0
<19> DW_AT_high_pc : 19
<1a> DW_AT_name : (indirect string, offset: 0): arch/x86/entry/entry.S
<1e> DW_AT_comp_dir : (indirect string, offset: 0x17): /root/git/linux
<22> DW_AT_producer : (indirect string, offset: 0x27): GNU AS 2.39.50
<26> DW_AT_language : 32769 (MIPS assembler)
<1><28>: Abbrev Number: 2 (DW_TAG_subprogram)
<29> DW_AT_name : (indirect string, offset: 0x36): entry_ibpb
<2d> DW_AT_external : 1
<2d> DW_AT_type : <0x37>
<2e> DW_AT_low_pc : 0
<36> DW_AT_high_pc : 19
<1><37>: Abbrev Number: 3 (DW_TAG_unspecified_type)
<1><38>: Abbrev Number: 0
15e20ce2324a:~/git/linux #
Which pahole -J barfs on:
15e20ce2324a:~/git/linux # pahole -J ./arch/x86/entry/entry.o
[1] UNKNOWN (anon) Unexpected kind for reference
15e20ce2324a:~/git/linux #
But if we ask it to exclude asm CUs (<26> DW_AT_language : 32769
(MIPS assembler)) then it ignores it, so this is a workaround.
15e20ce2324a:~/git/linux # pahole --lang_exclude asm -V -J ./arch/x86/entry/entry.o
15e20ce2324a:~/git/linux #
Now I'm looking at how to get the BTF encoder grokking this.
This is what I came up with, Andrii, Yonghong, wdyt?
Does `const void` make sense? Why not just keeping it as "void"?
"const void" might confuse tooling and BTF verifier in kernel, but I
haven't checked. Just trying to understand why we need extra "const".
I was trying to have an entry for that DW_TAG_unspecified_type type
instead of plain changing the type for entry_ibpb (a DW_TAG_subprogram)
straight to 0 (void).
I think this is the best way forward at the moment, i.e. to just change
the type for such DW_TAG_subprogram to zero, aka void, losing this
information by default.
I'll have a --btf_encode_unspecified_type as this const void trick as an
opt-in, who knows if this ends up being useful :-)
For function entry_ibpb, actually the kernel has a prorotype,
include/asm/nospec-branch.h:extern void entry_ibpb(void);
But unfortunately since the function is defined in asm code.
The actual type information is not encoded in dwarf so pahole
does not have enough info from vmlinux dwarf to get func
types. The compiler does not generate func types based on
declarations.
This may prevent people from tracing functions defined in
asm code but this should be extremely rare.
So I agree that BTF type id 0 is probably the best choice.
- Arnaldo
I.e. we wouldn't lose that info that the return type for that "function"
(it was encoded as a DW_TAG_subprogram) isn't defined. Using straight
void would state that it doesn't return anything.
I couldn't find a way to use KCFLAGS to ask Kbuild to pass that -B flag
to gcc so that I could rebuild the whole kernel with it to check if the
build goes thru and then try booting with such BTF info.
- Arnaldo
As an additional info clang 14 (haven't tested with other versions)
encodes such ASM Labels as DW_TAG_label and this thus isn't an issue
there.
- Arnaldo
commit 15ec614672da043008df31aa6ee85ebc5105d4fd
Author: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Date: Tue Oct 4 18:22:53 2022 -0300
btf_encoder: Encode DW_TAG_unspecified_type as BTF_KIND_CONST
This first appeared for assembler files in the Linux kernel with recent
GNU compilers, we don't have anything in BTF, AFAIK, to properly
represent that, so, for now, lets go with BTF_KIND_CONST.
Testing it:
Built binutils from git://sourceware.org/git/binutils-gdb.git, then used
gcc's -B option to point to the directory with the new as, that is built
as as-new, so make a symlink, ending up with:
15e20ce2324a:~/git/linux # readelf -wi ./arch/x86/entry/entry.o
Contents of the .debug_info section:
Compilation Unit @ offset 0:
Length: 0x35 (32-bit)
Version: 5
Unit Type: DW_UT_compile (1)
Abbrev Offset: 0
Pointer Size: 8
<0><c>: Abbrev Number: 1 (DW_TAG_compile_unit)
<d> DW_AT_stmt_list : 0
<11> DW_AT_low_pc : 0
<19> DW_AT_high_pc : 19
<1a> DW_AT_name : (indirect string, offset: 0): arch/x86/entry/entry.S
<1e> DW_AT_comp_dir : (indirect string, offset: 0x17): /root/git/linux
<22> DW_AT_producer : (indirect string, offset: 0x27): GNU AS 2.39.50
<26> DW_AT_language : 32769 (MIPS assembler)
<1><28>: Abbrev Number: 2 (DW_TAG_subprogram)
<29> DW_AT_name : (indirect string, offset: 0x36): entry_ibpb
<2d> DW_AT_external : 1
<2d> DW_AT_type : <0x37>
<2e> DW_AT_low_pc : 0
<36> DW_AT_high_pc : 19
<1><37>: Abbrev Number: 3 (DW_TAG_unspecified_type)
<1><38>: Abbrev Number: 0
So we have that asm label encoded by GNU AS 2.39.50 as a
DW_TAG_subprogram that has as its DW_AT_type the DW_TAG_unspecified_type
0x37 that we encode as a BTF_KIND_CONST pointing to 0 (void):
15e20ce2324a:~/git/linux # pahole -J ./arch/x86/entry/entry.o
15e20ce2324a:~/git/linux # pahole -JV ./arch/x86/entry/entry.o
btf_encoder__new: './arch/x86/entry/entry.o' doesn't have '.data..percpu' section
Found 0 per-CPU variables!
Found 1 functions!
File ./arch/x86/entry/entry.o:
[1] CONST (anon) type_id=0
[2] FUNC_PROTO (anon) return=1 args=(void)
[3] FUNC entry_ibpb type_id=2
15e20ce2324a:~/git/linux # pfunct -F btf ./arch/x86/entry/entry.o
entry_ibpb
15e20ce2324a:~/git/linux # pfunct --proto -F btf ./arch/x86/entry/entry.o
const void entry_ibpb(void);
15e20ce2324a:~/git/linux #
15e20ce2324a:~/git/linux # tools/bpf/bpftool/bpftool btf dump file ./arch/x86/entry/entry.o format raw
[1] CONST '(anon)' type_id=0
[2] FUNC_PROTO '(anon)' ret_type_id=1 vlen=0
[3] FUNC 'entry_ibpb' type_id=2 linkage=static
15e20ce2324a:~/git/linux #
I think this is what can be done to avoid having to skip ASM DWARF when
gets widely used, i.e. binutils gets updated.
Reported-by: Martin Liška <mliska@xxxxxxx>
Cc: Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx>
Cc: Yonghong Song <yhs@xxxxxx>
Signed-off-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
diff --git a/btf_encoder.c b/btf_encoder.c
index 7ad3f29ef153d8d6..7e50ba711ccc14d6 100644
--- a/btf_encoder.c
+++ b/btf_encoder.c
@@ -962,6 +962,8 @@ static int btf_encoder__encode_tag(struct btf_encoder *encoder, struct tag *tag,
return btf_encoder__add_enum_type(encoder, tag, conf_load);
case DW_TAG_subroutine_type:
return btf_encoder__add_func_proto(encoder, tag__ftype(tag), type_id_off);
+ case DW_TAG_unspecified_type:
+ return btf_encoder__add_ref_type(encoder, BTF_KIND_CONST, 0, NULL, false);
default:
fprintf(stderr, "Unsupported DW_TAG_%s(0x%x): type: 0x%x\n",
dwarf_tag_name(tag->tag), tag->tag, ref_type_id);
--
- Arnaldo
