Re: [PATCH 5/6] gdb: Remove cleanup from linux-fork.c:inferior_call_waitpid

[Pedro Alves <palves at redhat dot com>]

On 01/01/2019 10:45 PM, Andrew Burgess wrote:

> +/* Temporarily switch to the infrun state stored on the fork_info
> +   identified by a given ptid_t.  When this object goes out of scope,
> +   restore the currently selected infrun state.   */
> +
> +class scoped_switch_fork_info
>  {
> -  struct fork_info *oldfp = (struct fork_info *) fp;
> +public:
> +  /* Switch to the infrun state held on the fork_info identified by
> +     PPTID.  If PPTID is the current inferior then no switch is done.  */
> +  scoped_switch_fork_info (ptid_t pptid)

(Nit, "explicit scoped_switch_fork_info")

> +    : m_oldfp (nullptr)
> +  {
> +    if (pptid != inferior_ptid)
> +      {
> +	struct fork_info *newfp = nullptr;
> +
> +	/* Switch to pptid.  */
> +	m_oldfp = find_fork_ptid (inferior_ptid);
> +	gdb_assert (m_oldfp != nullptr);
> +	newfp = find_fork_ptid (pptid);
> +	gdb_assert (newfp != nullptr);
> +	fork_save_infrun_state (m_oldfp, 1);
> +	remove_breakpoints ();
> +	fork_load_infrun_state (newfp);
> +	insert_breakpoints ();
> +      }
> +  }
>  
> -  if (oldfp)
> -    {
> -      /* Switch back to inferior_ptid.  */
> -      remove_breakpoints ();
> -      fork_load_infrun_state (oldfp);
> -      insert_breakpoints ();
> -    }
> -}
> +  /* Restore the previously selected infrun state.  If the constructor
> +     didn't need to switch states, then nothing is done here either.  */
> +  ~scoped_switch_fork_info ()
> +  {
> +    if (m_oldfp != nullptr)
> +      {
> +	/* Switch back to inferior_ptid.  */
> +	remove_breakpoints ();
> +	fork_load_infrun_state (m_oldfp);
> +	insert_breakpoints ();
> +      }

When writing destructors, we need to keep in mind that if an
exception escapes, gdb is terminated on the spot.

Things aren't usually correct in the cleanup version either,
since an exception escaping while running cleanups leaves
the cleanup chain with dangling cleanups, but I think that
these conversions are the ideal time to fix it up.

The solution will usually be to swallow exceptions in the
dtor with try/catch(...) and try to limp along.

Thanks,
Pedro Alves