Re: [PATCH] gdb: use correct target in notify_thread_exited()

[Simon Marchi <simark@simark.ca>]

On 5/4/26 3:16 AM, Markus Metzger wrote:
> clean_up_just_stopped_threads_fsms() may call notify_thread_exited() in
> the context of a thread from a different target.  In my case:
> 
>     at .../gdb/thread.c:404
>     exit_code=std::optional [no contained value], silent=false)
>     at .../gdb/thread.c:436
>     exit_code=std::optional [no contained value], silent=false)
>     at .../gdb/thread.c:733
>     at .../gdb/thread.c:763
>     at .../gdb/infrun.c:4548
>     at .../gdb/infrun.c:4786

This is missing the function names, so not very useful.

> 
> Instead of relying on global state, use the exited thread's inferior's
> target, which is exactly the target we want.
> ---
>  gdb/thread.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/gdb/thread.c b/gdb/thread.c
> index 74124596b28..5c3b60b2aa7 100644
> --- a/gdb/thread.c
> +++ b/gdb/thread.c
> @@ -203,14 +203,16 @@ notify_thread_exited (thread_info *t, std::optional<ULONGEST> exit_code,
>  {
>    if (!silent && print_thread_events)
>      {
> +      struct target_ops *target = t->inf->top_target ();
> +
>        if (exit_code.has_value ())
>  	gdb_printf (_("[%s (id %s) exited with code %s]\n"),
> -		    target_pid_to_str (t->ptid).c_str (),
> +		    target->pid_to_str (t->ptid).c_str (),
>  		    print_thread_id (t),
>  		    pulongest (*exit_code));
>        else
>  	gdb_printf (_("[%s (id %s) exited]\n"),
> -		    target_pid_to_str (t->ptid).c_str (),
> +		    target->pid_to_str (t->ptid).c_str (),
>  		    print_thread_id (t));

This is unfortunately not correct, because target calls (annoyingly IMO)
use the target stack of the current inferior to find the target beneath.
If you are having to do this change, it means that this function is
called for a thread of inferior A, while inferior B is the current one.
What will happen is that you'll start with the top target of inferior A,
but if that target calls "target_ops::beneath()" to delegate to the
target beneath, that will jump to a target of inferior B.  If the
pid_to_str operation is handled by the top target of inferior A, it will
work, but not in the general case.

As of today, the only fix for that would be to temporarily switch
inferior.

I once made a prototype [1] of passing down the target stack explicitly to
all target functions.  It would be quite invasive, but I don't see any
other solution if we want to lift that "target calls care about current
inferior" limitation.

Simon

[1] https://review.lttng.org/c/binutils-gdb/+/5627