mbox series

[bpf-next,v3,00/13] bpf: add btf func info support

Message ID 20181017233026.1071519-1-yhs@fb.com
Headers show
Series bpf: add btf func info support | expand

Message

Yonghong Song Oct. 17, 2018, 11:30 p.m. UTC
The BTF support was added to kernel by Commit 69b693f0aefa
("bpf: btf: Introduce BPF Type Format (BTF)"), which introduced
.BTF section into ELF file and is primarily
used for map pretty print.
pahole is used to convert dwarf to BTF for ELF files.

This patch added func info support to the kernel so we can
get better ksym's for bpf function calls. Basically,
pairs of bpf function calls and their corresponding types
are passed to the kernel. Extracting function names from
the types, the kernel is able to construct a ksym for
each function call with embedded function name.

This patch set added support of FUNC and FUNC_PROTO types
in the kernel. LLVM patch https://reviews.llvm.org/D53261
can generate func info, encoded in .BTF.ext ELF section.
The following is an example to show FUNC and
FUNC_PROTO difference, compiled with the above LLVM patch
with Debug mode.

  -bash-4.2$ cat test.c
  int foo(int (*bar)(int)) { return bar(5); }
  -bash-4.2$ clang -target bpf -g -O2 -mllvm -debug-only=btf -c test.c
  Type Table:
  [1] FUNC name_off=1 info=0x0c000001 size/type=2
        param_type=3
  [2] INT name_off=11 info=0x01000000 size/type=4
        desc=0x01000020
  [3] PTR name_off=0 info=0x02000000 size/type=4
  [4] FUNC_PROTO name_off=0 info=0x0d000001 size/type=2
        param_type=2

  String Table:
  0 :
  1 : foo
  5 : .text
  11 : int
  15 : test.c
  22 : int foo(int (*bar)(int)) { return bar(5); }

  FuncInfo Table:
  sec_name_off=5
        insn_offset=<Omitted> type_id=1
  ...
  
In the above, type and string tables are in .BTF section and
the func info in .BTF.ext. The "<Omitted>" is the
insn offset which is not available during the dump time but
resolved during later compilation process.
Following the format specification at Patch #9 and examine the
raw data in .BTF.ext section, we have
  FuncInfo Table:
  sec_name_off=5
        insn_offset=0 type_id=1
The (insn_offset, type_id) can be passed to the kernel
so the kernel can find the func name and use it in the ksym.
Below is a demonstration from Patch #13.
  $ bpftool prog dump jited id 1
  int _dummy_tracepoint(struct dummy_tracepoint_args * ):
  bpf_prog_b07ccb89267cf242__dummy_tracepoint:
         0:   push   %rbp
         1:   mov    %rsp,%rbp
        ......
        3c:   add    $0x28,%rbp
        40:   leaveq
        41:   retq
  
  int test_long_fname_1(struct dummy_tracepoint_args * ):
  bpf_prog_2dcecc18072623fc_test_long_fname_1:
         0:   push   %rbp
         1:   mov    %rsp,%rbp
        ......
        3a:   add    $0x28,%rbp
        3e:   leaveq
        3f:   retq
  
  int test_long_fname_2(struct dummy_tracepoint_args * ):
  bpf_prog_89d64e4abf0f0126_test_long_fname_2:
         0:   push   %rbp
         1:   mov    %rsp,%rbp
        ......
        80:   add    $0x28,%rbp
        84:   leaveq
        85:   retq

For the patchset,
Patch #1  refactors the code to break up btf_type_is_void().
Patch #2  introduces new BTF types BTF_KIND_FUNC and BTF_KIND_FUNC_PROTO.
Patch #3  syncs btf.h header to tools directory.
Patch #4  adds btf func/func_proto self tests in test_btf.
Patch #5  adds kernel interface to load func_info to kernel
          and pass func_info to userspace.
Patch #6  syncs bpf.h header to tools directory.
Patch #7  adds news btf/func_info related fields in libbpf
          program load function.
Patch #8  extends selftest test_btf to test load/retrieve func_type info.
Patch #9  adds .BTF.ext func info support.
Patch #10 changes Makefile to avoid using pahole if llvm is capable of
          generating BTF sections.
Patch #11 refactors to have btf_get_from_id() in libbpf for reuse.
Patch #12 enhance test_btf file testing to test func info.
Patch #13 adds bpftool support for func signature dump.

Changelogs:
  v2 -> v3:
    . Removed kernel btf extern functions btf_type_id_func()
      and btf_get_name_by_id(). Instead, exposing existing
      functions btf_type_by_id() and btf_name_by_offset().
    . Added comments about ELF section .BTF.ext layout.
    . Better codes in btftool as suggested by Edward Cree.
  v1 -> v2:
    . Added missing sign-off.
    . Limited the func_name/struct_member_name length for validity test.
    . Removed/changed several verifier messages.
    . Modified several commit messages to remove line_off reference.

Yonghong Song (13):
  bpf: btf: Break up btf_type_is_void()
  bpf: btf: Add BTF_KIND_FUNC and BTF_KIND_FUNC_PROTO
  tools/bpf: sync kernel btf.h header
  tools/bpf: add btf func/func_proto unit tests in selftest test_btf
  bpf: get better bpf_prog ksyms based on btf func type_id
  tools/bpf: sync kernel uapi bpf.h header to tools directory
  tools/bpf: add new fields for program load in lib/bpf
  tools/bpf: extends test_btf to test load/retrieve func_type info
  tools/bpf: add support to read .BTF.ext sections
  tools/bpf: do not use pahole if clang/llvm can generate BTF sections
  tools/bpf: refactor to implement btf_get_from_id() in lib/bpf
  tools/bpf: enhance test_btf file testing to test func info
  tools/bpf: bpftool: add support for jited func types

 include/linux/bpf.h                          |   2 +
 include/linux/bpf_verifier.h                 |   1 +
 include/linux/btf.h                          |   2 +
 include/uapi/linux/bpf.h                     |  11 +
 include/uapi/linux/btf.h                     |   9 +-
 kernel/bpf/btf.c                             | 313 ++++++++--
 kernel/bpf/core.c                            |  13 +
 kernel/bpf/syscall.c                         |  86 ++-
 kernel/bpf/verifier.c                        |  46 ++
 samples/bpf/Makefile                         |   8 +
 tools/bpf/bpftool/btf_dumper.c               |  90 +++
 tools/bpf/bpftool/main.h                     |   2 +
 tools/bpf/bpftool/map.c                      |  68 +--
 tools/bpf/bpftool/prog.c                     |  54 ++
 tools/include/uapi/linux/bpf.h               |  11 +
 tools/include/uapi/linux/btf.h               |   9 +-
 tools/lib/bpf/bpf.c                          |   3 +
 tools/lib/bpf/bpf.h                          |   3 +
 tools/lib/bpf/btf.c                          | 305 ++++++++++
 tools/lib/bpf/btf.h                          |  50 ++
 tools/lib/bpf/libbpf.c                       |  53 +-
 tools/testing/selftests/bpf/Makefile         |   8 +
 tools/testing/selftests/bpf/test_btf.c       | 581 ++++++++++++++++++-
 tools/testing/selftests/bpf/test_btf_haskv.c |  16 +-
 tools/testing/selftests/bpf/test_btf_nokv.c  |  16 +-
 25 files changed, 1621 insertions(+), 139 deletions(-)

Comments

Alexei Starovoitov Oct. 18, 2018, 3:24 a.m. UTC | #1
On Wed, Oct 17, 2018 at 04:30:22PM -0700, Yonghong Song wrote:
> 
> This patch added func info support to the kernel so we can
> get better ksym's for bpf function calls. Basically,
> pairs of bpf function calls and their corresponding types
> are passed to the kernel. Extracting function names from
> the types, the kernel is able to construct a ksym for
> each function call with embedded function name.
...
> Below is a demonstration from Patch #13.
>   $ bpftool prog dump jited id 1
>   int _dummy_tracepoint(struct dummy_tracepoint_args * ):
>   bpf_prog_b07ccb89267cf242__dummy_tracepoint:
>          0:   push   %rbp
>          1:   mov    %rsp,%rbp
>         ......
>         3c:   add    $0x28,%rbp
>         40:   leaveq
>         41:   retq
>   
>   int test_long_fname_1(struct dummy_tracepoint_args * ):
>   bpf_prog_2dcecc18072623fc_test_long_fname_1:
>          0:   push   %rbp
>          1:   mov    %rsp,%rbp

Considering we only had 16-byte names for main prog and
no names at all for subprogs, this is huge improvement
in BPF introspection.
Cannot wait for the followup patches with line info support.

For the set:
Acked-by: Alexei Starovoitov <ast@kernel.org>