Use DTrace to diagnose gdb issues

A few days ago, I installed the newest 64-bit gdb program (version 7.7.1) on Solaris 10 (X86_64 platform) to debug programs. After playing with the gdb a day, I found 2 issues about gdb:
(1) The “set follow-fork-mode child” command doesn’t take effect. By default, after the parent process forks the child process, the gdb will track the parent process, so this command can make gdb begin to follow the child process. But this command works OK on Linux.
(2) The gdb can’t parse the 32-bit application core dump file. Per my understanding, the 64-bit gdb should parse both 32-bit and 64-bit core dump files.

I thought these are 2 bugs, so I reported them to the gdb organization :https://sourceware.org/bugzilla/show_bug.cgi?id=17044 and https://sourceware.org/bugzilla/show_bug.cgi?id=17045. But unfortunately, there was no any responses from the gdb organization. Suddenly I hit upon an idea: since I worked on Solaris, why not use DTrace? Yes, DTrace. DTrace is a very cool application which can trace application, and tell you the function flow of the application. So I started working on it at once!

(1) The “set follow-fork-mode child” command doesn’t take effect.
Firstly, I wrote a simple C test program:

#include <stdio.h>
#include <sys/types.h>
#include <unistd.h>

int main(void) {
    pid_t pid;

    pid = fork();
    if (pid < 0)
    {
        exit(1);
    }
    else if (pid > 0)
    {
        exit(0);
    }
    printf("hello world\n");
    return 0;
}

Then I wrote the DTrace script (debug_gdb.d) to diagnose gdb. Because I hadn’t read gdb source code before and I didn’t know the function flow of the gdb, I would like to know how the functions are called and executed.
The script was like the following, and it recorded every entry and return of the function:

#!/usr/sbin/dtrace -Fs
pid$target:gdb::entry,  
pid$target:gdb::return
{
}

The “-F” option’s effect is “Coalesce trace output by identifying function entry and return. Function entry probe reports are indented and their output is prefixed with ->. Function return probe reports are unindented and their output is prefixed with <- . ” so it can tell us how the gdb was executed.

I began to use gdb to debug the program:

bash-3.2# ./gdb -data-directory ./data-directory /data/nan/a
GNU gdb (GDB) 7.7.1
Copyright (C) 2014 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-pc-solaris2.10".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
<http://www.gnu.org/software/gdb/documentation/>.
For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from /data/nan/a...done.

Before executing “set follow-fork-mode child” command, I ran the DTrace script (the 4408 is the gdb process ID):

`./debug_gdb.d -p 4408 > a.txt`

Then I executed the “set follow-fork-mode child” command in gdb:

(gdb) set follow-fork-mode child

Oh, the output of DTrace script was very large, more than 900 lines, but I found the following function entries and returns:

  CPU FUNCTION        
  ......
  5                      -> do_set_command    
  5                        -> do_sfunc        
  5                          -> empty_sfunc   
  5                          <- empty_sfunc   
  5                        <- do_set_command
  ......

From the name of empty_sfunc, it seemed that gdb did nothing, right? I began to search the code in gdb:
a) The empty_sfunc is like this:

static void
empty_sfunc (char *args, int from_tty, struct cmd_list_element *c)
{
}

It really do nothing!

and it is called in add_set_or_show_cmd:

static struct cmd_list_element *
add_set_or_show_cmd (const char *name,
             enum cmd_types type,
             enum command_class class,
             var_types var_type,
             void *var,
             char *doc,
             struct cmd_list_element **list)
{
  struct cmd_list_element *c = add_cmd (name, class, NULL, doc, list);

  gdb_assert (type == set_cmd || type == show_cmd);
  c->type = type;
  c->var_type = var_type;
  c->var = var;
  /* This needs to be something besides NULL so that this isn't
     treated as a help class.  */
  set_cmd_sfunc (c, empty_sfunc);
  return c;
}

and the set_cmd_sfunc is like this:

void
set_cmd_sfunc (struct cmd_list_element *cmd, cmd_sfunc_ftype *sfunc)
{
  if (sfunc == NULL)
    cmd->func = NULL;
  else
    cmd->func = do_sfunc;
  cmd->function.sfunc = sfunc; /* Ok.  */
}

so from above, I knew the cmd‘s function.sfunc is set to empty_sfunc.
b) Then I searched “follow-fork-mode” in source code, it appears in add_setshow_enum_cmd function:

add_setshow_enum_cmd ("follow-fork-mode", class_run,
        follow_fork_mode_kind_names,
        &follow_fork_mode_string, _("\
Set debugger response to a program call of fork or vfork."), _("\
Show debugger response to a program call of fork or vfork."), _("\
A fork or vfork creates a new process.  follow-fork-mode can be:\n\
  parent  - the original process is debugged after a fork\n\
  child   - the new process is debugged after a fork\n\
The unfollowed process will continue to run.\n\
By default, the debugger will follow the parent process."),
            NULL,
            show_follow_fork_mode_string,
            &setlist, &showlist);

The fourth input parameter from bottom is NULL, and it is the set_func of the cmd. The add_setshow_enum_cmd is like this:

void
add_setshow_enum_cmd (const char *name,
              enum command_class class,
              const char *const *enumlist,
              const char **var,
              const char *set_doc,
              const char *show_doc,
              const char *help_doc,
              cmd_sfunc_ftype *set_func,
              show_value_ftype *show_func,
              struct cmd_list_element **set_list,
              struct cmd_list_element **show_list)
{
  struct cmd_list_element *c;

  add_setshow_cmd_full (name, class, var_enum, var,
            set_doc, show_doc, help_doc,
            set_func, show_func,
            set_list, show_list,
            &c, NULL);
  c->enums = enumlist;
}

add_setshow_enum_cmd calls add_setshow_cmd_full, and what does add_setshow_cmd_full do?

static void
add_setshow_cmd_full (const char *name,
              enum command_class class,
              var_types var_type, void *var,
              const char *set_doc, const char *show_doc,
              const char *help_doc,
              cmd_sfunc_ftype *set_func,
              show_value_ftype *show_func,
              struct cmd_list_element **set_list,
              struct cmd_list_element **show_list,
              struct cmd_list_element **set_result,
              struct cmd_list_element **show_result)
{
    ......
    set = add_set_or_show_cmd (name, set_cmd, class, var_type, var,
             full_set_doc, set_list);
    set->flags |= DOC_ALLOCATED;

    if (set_func != NULL)
        set_cmd_sfunc (set, set_func);
    ......
}

Form the above code analysis, I got the following conclusion: by default, the cmd‘s function.sfunc is empty_sfunc, and it would be set only if the set_func is not NULL. Because the default value of “follow-fork-mode” command’s set_func isNULL, this command wouldn’t take effect by default.

Another question, why does it work OK on Linux? After further investigating the code, I found the Linux has a customized function for supporting this command: linux_child_follow_fork.

So the conclusion is: by default, the gdb doesn’t support “set follow-fork-mode child” command, it needs the different platforms implement the function themselves.

(2) The gdb can’t parse the 32-bit application core dump file.
Firstly, I wrote a simple C program which can generate core dump file:

#include <stdio.h>

int main(void) {
        int *p = NULL;
        *p = 0;
        return 0;
}

Then, after generating core dump file, I would use gdb to analyze it. Usually, using gdb to debug core dump file command likes this:

gdb path/to/the/executable path/to/the/coredump

But because our DTrace script need work after gdb has run, I would use another method to start gdb:

bash-3.2# ./gdb -data-directory ./data-directory
GNU gdb (GDB) 7.7.1
Copyright (C) 2014 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-pc-solaris2.10".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
<http://www.gnu.org/software/gdb/documentation/>.
For help, type "help".
Type "apropos word" to search for commands related to "word".
(gdb) file /data/nan/a
Reading symbols from /data/nan/a...done.

Before loading the core dump file, I ran the DTrace script (the 4859 is the gdb process ID):

./debug_gdb.d -p 4859 > a.txt

Then I loaded core dump file in gdb, and gdb outputted:

(gdb) core-file /var/core/core.a.4856.1403588180
warning: Couldn't find general-purpose registers in core file.
[Thread debugging using libthread_db enabled]
[New Thread 1 (LWP 1)]
warning: Couldn't find general-purpose registers in core file.
warning: Couldn't find general-purpose registers in core file.
Error in re-setting breakpoint -21: PC register is not available
Error in re-setting breakpoint -22: PC register is not available
Error in re-setting breakpoint -23: PC register is not available
Error in re-setting breakpoint -24: PC register is not available
Error in re-setting breakpoint -25: PC register is not available
Error in re-setting breakpoint -26: PC register is not available
Error in re-setting breakpoint -27: PC register is not available
Error in re-setting breakpoint -28: PC register is not available
Core was generated by `./a'.
warning: Couldn't find general-purpose registers in core file.
#0  <unavailable> in ?? ()

But this time, the Dtrace ouput file was terribly large (almost 2G), and I couldn’t easily get the root cause like the above issue. But I noticed a function:warning. Yes, the gdb outputted “warning: Couldn’t find general-purpose registers in core file.”, so I hoped this function can help me, then I created a new DTrace script:

#!/usr/sbin/dtrace -Fs
pid$target:gdb:warning:entry,
pid$target:gdb:warning:return
{
        ustack();
}

When the warning was ouputted, the call stack would also be printted.

After executing “core-file /var/core/core.a.4856.1403588180” command again, the DTrace script output was like this:

10  -> warning                               
              gdb`warning
              gdb`get_core_register_section+0x180
              gdb`get_core_registers+0x16b
              gdb`sol_thread_fetch_registers+0x73
              gdb`target_fetch_registers+0x32
              gdb`ps_lgetfpregs+0x99
              libc_db.so.1`td_thr_getfpregs+0x4c
              gdb`sol_thread_fetch_registers+0x10d
              gdb`target_fetch_registers+0x32
              gdb`regcache_raw_read+0x123
              gdb`regcache_cooked_read_unsigned+0x87
              gdb`regcache_read_pc+0x58
              gdb`switch_to_thread+0x1d6
              gdb`switch_to_program_space_and_thread+0x7f
              gdb`breakpoint_re_set_one+0x65
              gdb`catch_errors+0x63
              gdb`breakpoint_re_set+0x76
              gdb`solib_add+0x190
              gdb`post_create_inferior+0xe1
              gdb`core_open+0x2a8

From this output, I could see an error occurred in get_core_register_section, then I read this function code:

static void
get_core_register_section (struct regcache *regcache,
               const char *name,
               int which,
               const char *human_name,
               int required)
{
 ......
  section = bfd_get_section_by_name (core_bfd, section_name);
  if (! section)
    {
      if (required)
    warning (_("Couldn't find %s registers in core file."),
         human_name);
      return;
    }
 ......
}

What does bfd_get_section_by_name do?

asection *
bfd_get_section_by_name (bfd *abfd, const char *name)
{
  struct section_hash_entry *sh;

  sh = section_hash_lookup (&abfd->section_htab, name, FALSE, FALSE);
  if (sh != NULL)
    return &sh->section;

  return NULL;
}

And the section_hash_lookup is like this:

#define section_hash_lookup(table, string, create, copy) \
  ((struct section_hash_entry *) \
   bfd_hash_lookup ((table), (string), (create), (copy)))

So until tracing here, I could see the root cause was bfd_hash_lookup return NULL, then gdb outputted “warning: Couldn’t find general-purpose registers in core file”.

I improved the DTrace script, and it was like this:

#!/usr/sbin/dtrace -Fs
pid$target:gdb:warning:entry
{
        printf("------------------BEGIN-------------------\n");
}
pid$target:gdb:warning:return
{
        printf("------------------END-------------------\n");
}
pid$target:gdb:bfd_hash_lookup:entry
{
        ustack();
        printf("%p, %s, %d, %d\n", arg0, copyinstr(arg1), arg2, arg3);
}
pid$target:gdb:bfd_hash_lookup:return
{
        ustack();
        printf("%p\n", arg1);
}

This script would print the input arguments and return values of the bfd_hash_lookup, so I can see why the warning was outputted.

After executing it, I knew the root cause was some register sections (like .reg) couldn’t be created if don’t exist.

As a man who haven’t read gdb source code, only through very simple DTrace scripts (contain only several lines), I can analyze the root cause of 2 gdb issues. So I suggest every Unix programmer try to learn DTrace, and it can give you big rewards!

Happy DTracing! Happy hacking!

One thought on “Use DTrace to diagnose gdb issues”

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.