Re: [PATCH v2 4/6] Python QUIT processing updates

[Pedro Alves <pedro@palves.net>]

On 2021-08-23 12:20 a.m., Kevin Buettner via Gdb-patches wrote:
> See the previous patch in this series for the motivation behind these
> updates.
> 
> This commit contains updates to Python's QUIT handling.  Specifically,
> it makes sure that a SIGTERM sent to GDB will be propagated to the top
> level.  While there are instances where we want to allow python code
> to catch Ctrl-C / SIGINT (and possibly) swallow it, I don't think we
> want that behavior for SIGTERM.  Therefore, there are cases where
> sync_quit_force_run is checked explicitly.  In other cases, I
> was able to defer the decision to GDB_PY_HANDLE_EXCEPTION.  (My goal
> was to minimize explicit references to sync_quit_force_run.)

Something to ponder about is whether having a distinct gdb_exception_sigterm
would be cleaner.

> ---
>  gdb/python/py-finishbreakpoint.c |  6 ++++--
>  gdb/python/py-gdb-readline.c     | 12 ++++++++----
>  gdb/python/py-symbol.c           |  4 ++++
>  gdb/python/py-utils.c            |  4 +++-
>  gdb/python/py-value.c            |  4 ++++
>  5 files changed, 23 insertions(+), 7 deletions(-)
> 
> diff --git a/gdb/python/py-finishbreakpoint.c b/gdb/python/py-finishbreakpoint.c
> index 1d8373d807e..120e40db6eb 100644
> --- a/gdb/python/py-finishbreakpoint.c
> +++ b/gdb/python/py-finishbreakpoint.c
> @@ -270,8 +270,10 @@ bpfinishpy_init (PyObject *self, PyObject *args, PyObject *kwargs)
>      }
>    catch (const gdb_exception &except)
>      {
> -      /* Just swallow.  Either the return type or the function value
> -	 remain NULL.  */
> +      /* Swallow everything except for a SIGTERM sent to GDB.  Either the
> +         return type or the function value remain NULL.  */
> +      if (sync_quit_force_run)
> +	throw;

AFAIK, it is not safe to let C++ exceptions cross the Python runtime, as Python
is written in C.  The issue may be papered over on x86-64 Linux and there
everything is built with -fasynchronous-unwind-tables.  Even if it looks like
it works, Python doesn't see the exception at all so isn't given a chance to
properly cleanup as the exception unwinds the stack.

The same concern applies to the rest of the patch.