Manohar Manohar - 3 months ago 6
Linux Question

gcc debug symbols (-g flag) vs linker's -rdynamic option

glibc provides

backtrace()
and
backtrace_symbols()
to get the stack trace of a running program. But for this to work the program has to be built with linker's
-rdynamic
flag.

What is the difference between
-g
flag passed to gcc vs linker's
-rdynamic
flag ? For a sample code I did readelf to compare the outputs.
-rdynamic
seems to produce more info under
Symbol table '.dynsym'
But I am not quite sure what the additional info is.

Even if I
strip
a program binary built using
-rdynamic
,
backtrace_symbols()
continue to work.

When
strip
removes all the symbols from the binary why is it leaving behind whatever was added by the
-rdynamic
flag ?

Edit: Follow-up questions based on Mat's response below..

For the same sample code you took this is the difference I see with
-g
&
-rdynamic


without any option..

Symbol table '.dynsym' contains 4 entries:
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000000000 218 FUNC GLOBAL DEFAULT UND __libc_start_main@GLIBC_2.2.5 (2)
2: 0000000000000000 0 NOTYPE WEAK DEFAULT UND _Jv_RegisterClasses
3: 0000000000000000 0 NOTYPE WEAK DEFAULT UND __gmon_start__

Symbol table '.symtab' contains 70 entries:
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000400200 0 SECTION LOCAL DEFAULT 1
2: 000000000040021c 0 SECTION LOCAL DEFAULT 2


with
-g
there are more sections, more entries in
.symtab
table but
.dynsym
remains the same..

[26] .debug_aranges PROGBITS 0000000000000000 0000095c
0000000000000030 0000000000000000 0 0 1
[27] .debug_pubnames PROGBITS 0000000000000000 0000098c
0000000000000023 0000000000000000 0 0 1
[28] .debug_info PROGBITS 0000000000000000 000009af
00000000000000a9 0000000000000000 0 0 1
[29] .debug_abbrev PROGBITS 0000000000000000 00000a58
0000000000000047 0000000000000000 0 0 1
[30] .debug_line PROGBITS 0000000000000000 00000a9f
0000000000000038 0000000000000000 0 0 1
[31] .debug_frame PROGBITS 0000000000000000 00000ad8
0000000000000058 0000000000000000 0 0 8
[32] .debug_loc PROGBITS 0000000000000000 00000b30
0000000000000098 0000000000000000 0 0 1

Symbol table '.dynsym' contains 4 entries:
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000000000 218 FUNC GLOBAL DEFAULT UND __libc_start_main@GLIBC_2.2.5 (2)
2: 0000000000000000 0 NOTYPE WEAK DEFAULT UND _Jv_RegisterClasses
3: 0000000000000000 0 NOTYPE WEAK DEFAULT UND __gmon_start__

Symbol table '.symtab' contains 77 entries:
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000400200 0 SECTION LOCAL DEFAULT 1


with
-rdynamic
no additional debug sections, .symtab entries are 70 (same as vanilla gcc invocation), but more
.dynsym
entries..

Symbol table '.dynsym' contains 19 entries:
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000000000 218 FUNC GLOBAL DEFAULT UND __libc_start_main@GLIBC_2.2.5 (2)
2: 00000000005008e8 0 OBJECT GLOBAL DEFAULT ABS _DYNAMIC
3: 0000000000400750 57 FUNC GLOBAL DEFAULT 12 __libc_csu_fini
4: 00000000004005e0 0 FUNC GLOBAL DEFAULT 10 _init
5: 0000000000400620 0 FUNC GLOBAL DEFAULT 12 _start
6: 00000000004006f0 86 FUNC GLOBAL DEFAULT 12 __libc_csu_init
7: 0000000000500ab8 0 NOTYPE GLOBAL DEFAULT ABS __bss_start
8: 00000000004006de 16 FUNC GLOBAL DEFAULT 12 main
9: 0000000000500aa0 0 NOTYPE WEAK DEFAULT 23 data_start
10: 00000000004007c8 0 FUNC GLOBAL DEFAULT 13 _fini
11: 00000000004006d8 6 FUNC GLOBAL DEFAULT 12 foo
12: 0000000000500ab8 0 NOTYPE GLOBAL DEFAULT ABS _edata
13: 0000000000500a80 0 OBJECT GLOBAL DEFAULT ABS _GLOBAL_OFFSET_TABLE_
14: 0000000000500ac0 0 NOTYPE GLOBAL DEFAULT ABS _end
15: 00000000004007d8 4 OBJECT GLOBAL DEFAULT 14 _IO_stdin_used
16: 0000000000500aa0 0 NOTYPE GLOBAL DEFAULT 23 __data_start
17: 0000000000000000 0 NOTYPE WEAK DEFAULT UND _Jv_RegisterClasses
18: 0000000000000000 0 NOTYPE WEAK DEFAULT UND __gmon_start__

Symbol table '.symtab' contains 70 entries:
Num: Value Size Type Bind Vis Ndx Name
0: 0000000000000000 0 NOTYPE LOCAL DEFAULT UND
1: 0000000000400200 0 SECTION LOCAL DEFAULT 1
2: 000000000040021c 0 SECTION LOCAL DEFAULT 2


Now these are the questions I have..


  1. In gdb you can do bt to get the bactrace. If that works with just
    -g
    why do we need
    -rdynamic
    for backtrace_symbols to work ?

  2. Comparing the additions to
    .symtab
    with
    -g
    & additions to
    .dynsym
    with
    -rdynamic
    they are not exactly the same.. Does either one provide better debugging info compared to the other ?
    FWIW, size of the output produced is like this: with -g > with -rdynamic > with neither option

  3. What exactly is the usage of .dynsym ? Is it all the symbols exported by this binary ? In that case why is foo going into .dynsym because we are not compiling the code as a library.

  4. If I link my code using all static libraries then -rdynamic is not needed for backtrace_symbols to work ?


Mat Mat
Answer

According to the docs:

This instructs the linker to add all symbols, not only used ones, to the dynamic symbol table.

Those are not debug symbols, they are dynamic linker symbols. Those are not removed by strip since it would (in most cases) break the executable - they are used by the runtime linker to do the final link stage of your executable.

Example:

$ cat t.c
void foo() {}
int main() { foo(); return 0; }

Compile and link without -rdynamic (and no optimizations, obviously)

$ gcc -O0 -o t t.c
$ readelf -s t

Symbol table '.dynsym' contains 3 entries:
   Num:    Value          Size Type    Bind   Vis      Ndx Name
     0: 0000000000000000     0 NOTYPE  LOCAL  DEFAULT  UND 
     1: 0000000000000000     0 FUNC    GLOBAL DEFAULT  UND __libc_start_main@GLIBC_2.2.5 (2)
     2: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND __gmon_start__

Symbol table '.symtab' contains 50 entries:
   Num:    Value          Size Type    Bind   Vis      Ndx Name
     0: 0000000000000000     0 NOTYPE  LOCAL  DEFAULT  UND 
     1: 0000000000400270     0 SECTION LOCAL  DEFAULT    1 
....
    27: 0000000000000000     0 FILE    LOCAL  DEFAULT  ABS t.c
    28: 0000000000600e14     0 NOTYPE  LOCAL  DEFAULT   18 __init_array_end
    29: 0000000000600e40     0 OBJECT  LOCAL  DEFAULT   21 _DYNAMIC

So the executable has a .symtab with everything. But notice that .dynsym doesn't mention foo at all - it has the bare essentials in there. This is not enough information for backtrace_symbols to work. It relies on the information present in that section to match code addresses with function names.

Now compile with -rdynamic:

$ gcc -O0 -o t t.c -rdynamic
$ readelf -s t

Symbol table '.dynsym' contains 17 entries:
   Num:    Value          Size Type    Bind   Vis      Ndx Name
     0: 0000000000000000     0 NOTYPE  LOCAL  DEFAULT  UND 
     1: 0000000000000000     0 FUNC    GLOBAL DEFAULT  UND __libc_start_main@GLIBC_2.2.5 (2)
     2: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND __gmon_start__
     3: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND _Jv_RegisterClasses
     4: 0000000000601018     0 NOTYPE  GLOBAL DEFAULT  ABS _edata
     5: 0000000000601008     0 NOTYPE  GLOBAL DEFAULT   24 __data_start
     6: 0000000000400734     6 FUNC    GLOBAL DEFAULT   13 foo
     7: 0000000000601028     0 NOTYPE  GLOBAL DEFAULT  ABS _end
     8: 0000000000601008     0 NOTYPE  WEAK   DEFAULT   24 data_start
     9: 0000000000400838     4 OBJECT  GLOBAL DEFAULT   15 _IO_stdin_used
    10: 0000000000400750   136 FUNC    GLOBAL DEFAULT   13 __libc_csu_init
    11: 0000000000400650     0 FUNC    GLOBAL DEFAULT   13 _start
    12: 0000000000601018     0 NOTYPE  GLOBAL DEFAULT  ABS __bss_start
    13: 000000000040073a    16 FUNC    GLOBAL DEFAULT   13 main
    14: 0000000000400618     0 FUNC    GLOBAL DEFAULT   11 _init
    15: 00000000004007e0     2 FUNC    GLOBAL DEFAULT   13 __libc_csu_fini
    16: 0000000000400828     0 FUNC    GLOBAL DEFAULT   14 _fini

Symbol table '.symtab' contains 50 entries:
   Num:    Value          Size Type    Bind   Vis      Ndx Name
     0: 0000000000000000     0 NOTYPE  LOCAL  DEFAULT  UND 
     1: 0000000000400270     0 SECTION LOCAL  DEFAULT    1 
....
    27: 0000000000000000     0 FILE    LOCAL  DEFAULT  ABS t.c
    28: 0000000000600e14     0 NOTYPE  LOCAL  DEFAULT   18 __init_array_end
    29: 0000000000600e40     0 OBJECT  LOCAL  DEFAULT   21 _DYNAMIC

Same thing for symbols in .symtab, but now foo has a symbol in the dynamic symbol section (and a bunch of other symbols appear there now too). This makes backtrace_symbols work - it now has enough information (in most cases) to map code addresses with function names.

Strip that:

$ strip --strip-all t
$ readelf -s t

Symbol table '.dynsym' contains 17 entries:
   Num:    Value          Size Type    Bind   Vis      Ndx Name
     0: 0000000000000000     0 NOTYPE  LOCAL  DEFAULT  UND 
     1: 0000000000000000     0 FUNC    GLOBAL DEFAULT  UND __libc_start_main@GLIBC_2.2.5 (2)
     2: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND __gmon_start__
     3: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND _Jv_RegisterClasses
     4: 0000000000601018     0 NOTYPE  GLOBAL DEFAULT  ABS _edata
     5: 0000000000601008     0 NOTYPE  GLOBAL DEFAULT   24 __data_start
     6: 0000000000400734     6 FUNC    GLOBAL DEFAULT   13 foo
     7: 0000000000601028     0 NOTYPE  GLOBAL DEFAULT  ABS _end
     8: 0000000000601008     0 NOTYPE  WEAK   DEFAULT   24 data_start
     9: 0000000000400838     4 OBJECT  GLOBAL DEFAULT   15 _IO_stdin_used
    10: 0000000000400750   136 FUNC    GLOBAL DEFAULT   13 __libc_csu_init
    11: 0000000000400650     0 FUNC    GLOBAL DEFAULT   13 _start
    12: 0000000000601018     0 NOTYPE  GLOBAL DEFAULT  ABS __bss_start
    13: 000000000040073a    16 FUNC    GLOBAL DEFAULT   13 main
    14: 0000000000400618     0 FUNC    GLOBAL DEFAULT   11 _init
    15: 00000000004007e0     2 FUNC    GLOBAL DEFAULT   13 __libc_csu_fini
    16: 0000000000400828     0 FUNC    GLOBAL DEFAULT   14 _fini
$ ./t
$

No .symtab is gone, but the dynamic symbol table is still there, and the executable runs. So backtrace_symbols still works too.

Strip the dynamic symbol table:

$ strip -R .dynsym t
$ ./t
./t: relocation error: ./t: symbol , version GLIBC_2.2.5 not defined in file libc.so.6 with link time reference

... and you get a broken executable.

An interesting read for what .symtab and .dynsym are used for is here: Inside ELF Symbol Tables. One of the things to note is that .symtab is not needed at runtime, so it is discarded by the loader. That section does not remain in the process's memory. .dynsym, on the otherhand, is needed at runtime, so it is kept in the process image. So it is available for things like backtrace_symbols to gather information about the current process from within itself.

So in short:

  • dynamic symbols are not stripped by strip since that would render the executable non-loadable
  • backtrace_symbols needs dynamic symbols to figure out what code belongs which function
  • backtrace_symbols does not use debugging symbols

Hence the behavior you noticed.


For your specific questions:

  1. gdb is a debugger. It uses debug information in the executable and libraries to display relevant information. It is much more complex than backtrace_symbols, and inspects the actual files on your drive in addition to the live process. backtrace_symbols does not, it is entirely in-process - so it cannot access sections that are not loaded into the executable image. Debug sections are not loaded into the runtime image, so it can't use them.
  2. .dynsym is not a debugging section. It is a section used by the dynamic linker. .symbtab isn't a debugging section either, but it can be used by debugger that have access to the executable (and library) files. -rdynamic does not generate debug sections, only that extended dynamic symbol table. The executable growth from -rdynamic depends entirely on the number of symbols in that executable (and alignment/padding considerations). It should be considerably less than -g.
  3. Except for statically linked binaries, executables need external dependencies resolved at load time. Like linking printf and some application startup procedures from the C library. These external symbols must be indicated somewhere in the executable: this is what .dynsym is used for, and this is why the exe has a .dynsym even if you don't specify -rdynamic. When you do specify it, the linker adds other symbols that are not necessary for the process to work, but can be used by things like backtrace_symbols.
  4. backtrace_symbols will not resolve any function names if you statically link. Even if you specify -rdynamic, the .dynsym section will not be emitted to the executable. No symbol tables gets loaded into the executable image, so backtrace_symbols cannot map code adresses to symbols.