[PATCH 4/4] Cygwin: pty: Guard accept_input routing and flush stale readahead in fast path

[Johannes Schindelin]

Hi Takashi,

On Wed, 18 Mar 2026, Takashi Yano wrote:

> On Mon, 02 Mar 2026 14:24:40 +0000
> "Johannes Schindelin wrote:
> > From: Johannes Schindelin <johannes.schindelin@gmx.de>
> > 
> > diff --git a/winsup/cygwin/fhandler/pty.cc b/winsup/cygwin/fhandler/pty.cc
> > index dd7ea9038..fcff53d88 100644
> > --- a/winsup/cygwin/fhandler/pty.cc
> > +++ b/winsup/cygwin/fhandler/pty.cc
> > @@ -489,6 +489,7 @@ fhandler_pty_master::accept_input ()
> >    HANDLE write_to = get_output_handle ();
> >    tmp_pathbuf tp;
> >    if (to_be_read_from_nat_pipe ()
> > +      && !get_ttyp ()->pcon_activated
> >        && get_ttyp ()->pty_input_state == tty::to_nat)
> >      {
> >        /* This code is reached if non-cygwin app is foreground and
> > @@ -2208,8 +2209,18 @@ fhandler_pty_master::write (const void *ptr, size_t len)
> >    WaitForSingleObject (input_mutex, mutex_timeout);
> >    if (to_be_read_from_nat_pipe () && get_ttyp ()->pcon_activated
> >        && get_ttyp ()->pty_input_state == tty::to_nat)
> > -    { /* Reaches here when non-cygwin app is foreground and pseudo console
> > -	 is activated. */
> > +    {
> > +      /* Flush any stale readahead data from a prior line_edit call that
> > +	 ran while pty_input_state was temporarily to_cyg (e.g. during a
> > +	 setpgid_aux transition when a cygwin child of the native process
> > +	 started or exited).  Without this, the readahead contents would
> > +	 be stranded and emitted after the direct WriteFile below,
> > +	 breaking chronological order. */
> > +      if (get_readahead_valid ())
> > +	{
> > +	  accept_input ();
> 
> Does the code path really reach here?
> At the end of pcon_start phase, accept_input() is already called as
> shown below. After that, transfer_input (tty::to_nat, ...) is called.
> Therefore, in the pcon_activated state, all key input will go to nat-
> pipe, not to readahead-buffer/cyg-pipe.
> 
>       if (!get_ttyp ()->pcon_start)
>         { /* Pseudo console initialization has been done in above code. */
>           pinfo pp (get_ttyp ()->pcon_start_pid);
>           if (get_ttyp ()->switch_to_nat_pipe
>               && pp && pp->pgid == get_ttyp ()->getpgid ()
>               && get_ttyp ()->pty_input_state_eq (tty::to_cyg))
>             {
>               /* This accept_input() call is needed in order to transfer input
>                  which is not accepted yet to non-cygwin pipe. */
>               WaitForSingleObject (input_mutex, mutex_timeout);
>               if (get_readahead_valid ())
>                 accept_input ();                   // <========== This
>               acquire_attach_mutex (mutex_timeout);
>               fhandler_pty_slave::transfer_input (tty::to_nat, from_master,
>                                                   get_ttyp (),
>                                                   input_available_event);
>               release_attach_mutex ();
>               ReleaseMutex (input_mutex);
>             }
>           get_ttyp ()->pcon_start_pid = 0;
>         }
> 
> What situation do you assume? Is there any case that the key input goes
> into cyg-pipe during pcon_activated other than mask_switch_to_nat_pipe()
> case?

In my AutoHotKey-based tests, this patch was required to fix the keystroke
out-of-order delivery.

However, adding `!pcon_activated` broke the routing in the `disable_pcon
case`. The readahead flush was addressing a symptom of the oscillation
rather than the root cause.

In your v7 series, `accept_input()` routes to the `nat` pipe whenever
`to_be_read_from_nat_pipe() && pty_input_state == to_nat`, regardless of
`pcon_activated`, which does not have the shortcomings of this here patch.

The readahead flush I added in this patch is handled differently (and
better) by your v7 4/7 via the `input_transferred_to_cyg`.

In short: I retract all of my patches, and thank you profusely for taking
the time to review them.

Ciao,
Johannes