GDB: problem debugging 32 bit binary on 64 bit machine

GDB: problem debugging 32 bit binary on 64 bit machine

[Jeff Kenton]

A little convoluted, but here's the story:

On our 64 machine gdb fails to read share libraries when debugging 32 
bit binaries (i.e., compiled with "-m32").  Debugging code built 
"-static" is OK.  The error is "Cannot access memory at address 
0x400000008" (address is above the 32 bit limit).  This is caused by a 
call to extract_typed_address() from scan_dyntag() reading an 8 byte 
type when it should only be reading 4 bytes.

So, I tweaked extract_typed_address() to know it was dealing with 32 
binaries.  Now gdb starts happily and reads in the shared libraries but 
the program won't execute.  It gets

warning: Can't read pathname for load map: Input/output error.
Cannot access memory at address 0x35dc3000

and the backtrace shows that it's in dl_main():
#0  *__GI__dl_debug_state () at dl-debug.c:77
#1  0x0000000077f93d00 in dl_main () at rtld.c:1651
#2  0x0000000077fb1698 in _dl_sysdep_start ()  at ../elf/dl-sysdep.c:244
#3  0x0000000077f90c70 in _dl_start_final () at rtld.c:334
#4  0x0000000077f96630 in _dl_start () at rtld.c:562
#5  0x0000000077fb2530 in _start () from .../lib32/ld.so.1

I have two questions:
1. is tweaking extract_typed_address() the right way to handle 32 bit 
addresses on a 64 bit machine? It seems weird but nothing else came to mind.
2. what's going wrong with dl_main()?
3. meta-question: is this a generic bug in gdb's handling of 32 bit 
binaries, or is there likely something I missed while porting?

Thanks.

--jeff

Re: GDB: problem debugging 32 bit binary on 64 bit machine

[Jan Kratochvil]

On Tue, 29 Nov 2011 21:58:14 +0100, Jeff Kenton wrote:
> 0x400000008" (address is above the 32 bit limit).  This is caused by
> a call to extract_typed_address() from scan_dyntag() reading an 8
> byte type when it should only be reading 4 bytes.

ARCH_SIZE there should be 32 and TARGET_GDBARCH should be 32-bit.

I guess for some reasons your GDB found wrong (64-bit) library for your 32-bit
program.  See the settings like `set solib-search-path', `set sysroot' etc.

> I have two questions:
> 1. is tweaking extract_typed_address() the right way to handle 32
> bit addresses on a 64 bit machine?

No.  If ARCH_SIZE and TARGET_GDBARCH are set right it will work.


> 2. what's going wrong with dl_main()?

There is called a notification new library has been loaded in the inferior,
therefore GDB tries to load a matching symbol file (=the library itself) on
the GDB side.


> 3. meta-question: is this a generic bug in gdb's handling of 32 bit
> binaries, or is there likely something I missed while porting?

It 32-on-64 normally works, at least in Fedora, no custom patches for it
there.


Regards,
Jan

Re: GDB: problem debugging 32 bit binary on 64 bit machine

[Jeff Kenton]

On 11/29/2011 04:11 PM, Jan Kratochvil wrote:
> On Tue, 29 Nov 2011 21:58:14 +0100, Jeff Kenton wrote:
>> 0x400000008" (address is above the 32 bit limit).  This is caused by
>> a call to extract_typed_address() from scan_dyntag() reading an 8
>> byte type when it should only be reading 4 bytes.
> ARCH_SIZE there should be 32 and TARGET_GDBARCH should be 32-bit.

This sounds like the basic problem.  Building the 32 bit target_gdbarch 
is easy.  What's the best way (place in the code) to detect 32 bit vs. 
64 bit binaries and switch between different gdbarch's?

Thanks.


>
> I guess for some reasons your GDB found wrong (64-bit) library for your 32-bit
> program.  See the settings like `set solib-search-path', `set sysroot' etc.
>
>> I have two questions:
>> 1. is tweaking extract_typed_address() the right way to handle 32
>> bit addresses on a 64 bit machine?
> No.  If ARCH_SIZE and TARGET_GDBARCH are set right it will work.
>
>
>> 2. what's going wrong with dl_main()?
> There is called a notification new library has been loaded in the inferior,
> therefore GDB tries to load a matching symbol file (=the library itself) on
> the GDB side.
>
>
>> 3. meta-question: is this a generic bug in gdb's handling of 32 bit
>> binaries, or is there likely something I missed while porting?
> It 32-on-64 normally works, at least in Fedora, no custom patches for it
> there.
>
>
> Regards,
> Jan

Re: GDB: problem debugging 32 bit binary on 64 bit machine

[Jan Kratochvil]

On Wed, 30 Nov 2011 14:35:14 +0100, Jeff Kenton wrote:
> This sounds like the basic problem.  Building the 32 bit
> target_gdbarch is easy.  What's the best way (place in the code) to
> detect 32 bit vs. 64 bit binaries and switch between different
> gdbarch's?

I do not understand the context.  The current GDB switches target_gdbarch
right during execution of "./gdb ./gdb", at:

#0  deprecated_target_gdbarch_select_hack (new_gdbarch=0x2071110) at gdbarch.c:4294
#1  in set_gdbarch_from_file (abfd=0x206c990) at arch-utils.c:559
#2  in exec_file_attach (filename=0x7fffffffde00 "./gdb", from_tty=1) at exec.c:294
#3  in catch_command_errors (command=0x488ab7 <exec_file_attach>, arg=0x7fffffffde00 "./gdb", from_tty=1, mask=6) at exceptions.c:531
#4  in captured_main (data=0x7fffffffd9e0) at ./main.c:850
#5  in catch_errors (func=0x4873f6 <captured_main>, func_args=0x7fffffffd9e0, errstring=0xe6ede7 "", mask=6) at exceptions.c:504
#6  in gdb_main (args=0x7fffffffd9e0) at ./main.c:953
#7  in main (argc=2, argv=0x7fffffffdae8) at gdb.c:35


Regards,
Jan