_start(ldso) & _start(crt)

_start(ldso) & _start(crt)

[writeonce]

Greetings,

When running a dynamically linked program, the general control flow is

     _start(ldso/{arch}/start.s) --> __dynlink --> _start(crt) --> 
__libc_start_main --> main()

When asking gdb to set up a breakpoint at _start (b _start), gdb's 
confirmation message shows the address of _start(crt), but it then 
breaks at _start(ldso), which is actually what I want.  It then later 
also breaks at _start(crt).

If gdb decided to only break at the executable's _start, then I believe 
there would be no way to request a break at the interpreter's (ldso)  
_start routine, since the address of ldso's _start is not known prior to 
execution.

I am not sure whether this is a good-enough motivation to rename ldso's 
start to _ldso_start (and set libc.so's entry point accordingly), or 
whether we could simply count on gdb and other debuggers to continue to 
do the "right thing" by breaking at the earlier _start of the two.  For 
one thing, setting a breakpoint at __dynlink provides pretty much the 
same debugging ability as with _start(ldso), and does not require any 
changes.  Any thoughts?

Kind regards,
zg

Re: _start(ldso) & _start(crt)

[Rich Felker]

On Mon, Apr 21, 2014 at 03:41:23PM -0400, writeonce@midipix.org wrote:
> Greetings,
> 
> When running a dynamically linked program, the general control flow is
> 
>     _start(ldso/{arch}/start.s) --> __dynlink --> _start(crt) -->
> __libc_start_main --> main()
> 
> When asking gdb to set up a breakpoint at _start (b _start), gdb's
> confirmation message shows the address of _start(crt), but it then
> breaks at _start(ldso), which is actually what I want.  It then
> later also breaks at _start(crt).
> 
> If gdb decided to only break at the executable's _start, then I
> believe there would be no way to request a break at the
> interpreter's (ldso)  _start routine, since the address of ldso's
> _start is not known prior to execution.
> 
> I am not sure whether this is a good-enough motivation to rename
> ldso's start to _ldso_start (and set libc.so's entry point
> accordingly), or whether we could simply count on gdb and other
> debuggers to continue to do the "right thing" by breaking at the
> earlier _start of the two.  For one thing, setting a breakpoint at
> __dynlink provides pretty much the same debugging ability as with
> _start(ldso), and does not require any changes.  Any thoughts?

The motivation for the name _start in the dynamic linker was
presumably that it gets used by default as the entry point, but we
already have -Wl,-e,_start anyway, so I don't see any compelling
reason it couldn't be renamed. Does anyone else object to changing it?

Rich

Re: _start(ldso) & _start(crt)

[M Farkas-Dyck]

Static-linked programs would be unaffected, yes?

Re: _start(ldso) & _start(crt)

[Rich Felker]

On Mon, Apr 21, 2014 at 04:15:25PM -0500, M Farkas-Dyck wrote:
> Static-linked programs would be unaffected, yes?

Indeed, this has nothing to do with static linked programs.

Rich