Re: Ensure correct symbol-file when attaching to a (remote) process

[Raphael Zulliger <zulliger@indel.ch>]

On 01/29/2013 04:18 AM, Raphael Zulliger wrote:
> I agree.
> (--build-id is a very interesting feature which I was not aware of. 
> Thanks for that hint Jan)
>
> In our scenario, our GDB stub could get that build-id from the running 
> target: Our embedded systems provides a variable read/write mechanism 
> accessible by the stub. Moreover, the embedded system could be made 
> aware of the address of the build-id by introducing variables around 
> the .note.gnu.build-id section in the linker script.
>
Here's the solution we came up with. Unfortunately, the solution is not 
very GDB-like mainly because the lack of time to create high quality GDB 
patches. But at least I want to share our (low-end) solution with you. 
(Note: I'm definitely not saying that this is the ideal solution! It's 
just good enough for our use cases.)

I post this message because of two purposes:
  1. As a reference if someone is searching the web
  2. To check if someone comes up with a better solution for reading the 
build-id out from an elf file

Here's what we've done in order to read the build-id from within our 
C/C++ code of our embedded software:

  - We pass "-Wl,--build-id" to the linker invocation

  - We extended the linker script by:
     .note    : {
         PROVIDE (__NOTE_BUILDID_BEGIN__ = .);
         *(.note.gnu.build-id)
         PROVIDE (__NOTE_BUILDID_END__ = .);
         } > ram

  - In the C/C++ code we have something like this:
   extern unsigned int __NOTE_BUILDID_BEGIN__;
   extern unsigned int __NOTE_BUILDID_END__;
With these two variables at hand, we are able to propagate the build-id 
to our GDB-stub by our own "PC to embedded system communication".

  - On the GDB side, we hacked some python code to read the build-id out 
from the ELF file:
def GetBuildId():
     BuildId = ""
     # we only ever have 1 ELF file (or 0 in case of a user error):
     for file in gdb.objfiles():
         cmd = 'objdump -s -j .note ' + file.filename
         p = subprocess.Popen(cmd, shell=True, stdin=subprocess.PIPE, 
stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
         output = p.stdout.read()
         for line in output.splitlines():
             if len(line) > 0 and line[0] == ' ':
                 for entry in line.split()[1:5]:
                     BuildId = BuildId + entry
     return BuildId

This is definitely no high-end code, but it does its job. Among others, 
It's potentially unstable if objdump changes it's output format... I 
searched the web for a better way of getting the content of a certain 
section out from an object file, but couldn't find a solution. If 
someone knows about a GDB/Python interface for doing so, I would 
definitely like to hear about it!

The build-id read out from the ELF file will then be sent to our GDB 
stub in a non-GDB way (again some kind of home-made communication, which 
we'd already in use). Our GDB stub finally compares the build-id read 
from the ELF file with the one read from the embedded system and can 
compare them for equality.

Raphael