[patch] gdbserver: Fix harmless crash for current-lwp-dead.exp

[Pedro Alves]

On Thursday 08 December 2011 00:05:00, Jan Kratochvil wrote:
> On Fri, 02 Dec 2011 23:10:43 +0100, Jan Kratochvil wrote:
> > gdb.threads/current-lwp-dead.exp with local/native gdbserver will make
> > crash/coredump of gdbserver on its shutdown.
> [...]
> > It is because gdbserver wants to kill the thread group leader - but in this
> > case there is no thread group leader, that is the purpose of the testcase.
> 
> I was wrong here, the TID == PID LWP is not a thread group leader.  It is from
> clone syscall without CLONE_THREAD, therefore with separate PID/TID.  It just
> belongs to the same gdbserver `struct process' as it has been caught by
> PTRACE_O_TRACECLONE.
> 
> As it is not a thread group leader waitpid already completed for it without
> any zombie LWP hanging, delete_lwp has been called.  This is the difference
> from a thread group leader - that despite it has PID == TID the waitpid has
> already completed for it.

Ah.

> > Hmm, do we leave the process zombie then?  Should we always
> > waitpid it to reap it?
> 
> There is nothing left from the kernel point of view.

When we fix gdbserver so that gdb.threads/leader-exit.exp doesn't
hang, then there will be.  On native debugging, we end up reapping
the zombie leader in inf-ptrace.c:inf_ptrace_mourn_inferior.  gdbserver
does not do that.

> Sorry for the confusion with "thread group leader".  OK to check it in?

Yes.

-- 
Pedro Alves