Re: [RFA/commit] GDB crash re-running program on Windows (native)

[Simon Marchi <simon dot marchi at polymtl dot ca>]

On 2018-12-30 22:57, Joel Brobecker wrote:
> Hello,
>
> Running any program twice on Windows current results in GDB crashing:
>
>     $ gdb -q any_program
>     (gdb) run
>     $ gdb dummy -batch -ex run -ex run
>     [New Thread 684960.0xe5878]
>     [New Thread 684960.0xd75ac]
>     [New Thread 684960.0xddac8]
>     [New Thread 684960.0xc1f50]
>     [Thread 684960.0xd75ac exited with code 0]
>     [Thread 684960.0xddac8 exited with code 0]
>     [Thread 684960.0xc1f50 exited with code 0]
>     [Inferior 1 (process 684960) exited normally]
>     (gdb) run
>     Segmentation fault
>
> The crash happens while processing the CREATE_PROCESS_DEBUG_EVENT
> for  the second run; in particular, we have in get_windows_debug_event:
>
>     | case CREATE_PROCESS_DEBUG_EVENT:
>     |   [...]
>     |   if (main_thread_id)
>     |     windows_delete_thread (ptid_t (current_event.dwProcessId, 0,
>     |                                    main_thread_id),
>     |                            0);
>
> The problem is that main_thread_id is the TID of the main thread from
> the *previous* inferior, and this code is trying to delete that
> thread. The problem is that it is constructing a PTID by pairing
> the TID of the previous inferior with the PID of the new inferior.
> As a result, when we dig inside windows_delete_thread to see
> how it would handle that, we see...
>
>     | delete_thread (find_thread_ptid (ptid));
>
> Since the PTID is bogus, we end up calling delete_thread with
> a NULL thread_info. It used to be harmless, turning the delete_thread
> into a nop, but the following change...
>
>     | commit 080363310650c93ad8e93018bcb6760ba5d32d1c
>     | Date:   Thu Nov 22 16:09:14 2018 +0000
>     | Subject: Per-inferior thread list, thread ranges/iterators, down
> with ALL_THREADS, etc.
>
> ... changed delete_thread to get the list of threads from
> the inferior, which itself is now accessed via the given
> thread_info. This is the corresponding diff that shows the change:
>
>     | -  for (tp = thread_list; tp; tpprev = tp, tp = tp->next)
>     | +  for (tp = thr->inf->thread_list; tp; tpprev = tp, tp = 
> tp->next)
>
> As a result of this, passing a NULL thread_info is no longer
> an option!
>
> Stepping back a bit, the reason behind deleting the thread late
> could be found in a patch from Dec 2003, which laconically explains:
>
>     | commit 87a45c96062d658ca83b50aa060a648bf5f5f1ff
>     | Date:   Fri Dec 26 00:39:04 2003 +0000
>     |
>     | * win32-nat.c (get_child_debug_event): Keep main thread id around
>     | even after thread exits since Windows insists on continuing to
>     | report events against it.
>
> A look at the gdb-patches archives did not provide any additional
> clues 
> (https://www.sourceware.org/ml/gdb-patches/2003-12/msg00478.html).
> It is not clear whether this is still needed or not. This patch
> assumes that whatever isue there was, the versions of Windows
> we currently support no longer have it.

This seems reasonable to me, if the testsuite shows no regression with 
the hack removed, I'm confident enough that this is ok.

> With that in mind, this commit fixes the issue by deleting the thread
> when the inferior sends the exit-process event as opposed to deleting 
> it
> later, while starting a new inferior.
>
> This restores also restores the printing of the thread-exit 
> notification

"This restores also restores"

> for the main thread, which was missing before. Looking at the 
> transcript
> of the example show above, we can see 4 thread creation notifications,
> and only 3 notifications for thread exits. Now creation and exit
> notifications are balanced.

Another choice is to not show the main thread's creation and exit (as is 
done on Linux), since it's kind of redundant with the process creation 
and exit.

> In the handling of EXIT_THREAD_DEBUG_EVENT, the main_thread_id
> check is removed because deemed unnecessary: The main thread was
> introduced by a CREATE_THREAD_DEBUG_EVENT, and thus the kernel
> is expected to report its death via EXIT_PROCESS_DEBUG_EVENT.

Should that last EXIT_PROCESS_DEBUG_EVENT actually be 
EXIT_THREAD_DEBUG_EVENT?

> @@ -1607,6 +1599,9 @@ get_windows_debug_event (struct target_ops *ops,
>  	}
>        else if (saw_create == 1)
>  	{
> +	  windows_delete_thread (ptid_t (current_event.dwProcessId, 0,
> +					 main_thread_id),
> +				 0);
>  	  ourstatus->kind = TARGET_WAITKIND_EXITED;
>  	  ourstatus->value.integer = current_event.u.ExitProcess.dwExitCode;
>  	  thread_id = main_thread_id;

If what you said above is right (that the kernel reports the main 
thread's death through an EXIT_THREAD_DEBUG_EVENT), why is this new call 
to windows_delete_thread needed?  Shouldn't it already be deleted at 
this point?

Simon