Re: [PATCH 2/2] tty: Implement lookahead to process XON/XOFF timely

From: Andy Shevchenko
Date: Wed Apr 06 2022 - 08:53:42 EST


On Wed, Apr 06, 2022 at 10:21:12AM +0200, Jiri Slaby wrote:
> On 05. 04. 22, 18:11, Andy Shevchenko wrote:
> > On Tue, Apr 05, 2022 at 01:24:37PM +0300, Ilpo Järvinen wrote:
> > > When tty is not read from, XON/XOFF may get stuck into an
> > > intermediate buffer. As those characters are there to do software
> > > flow-control, it is not very useful. In the case where neither end
> > > reads from ttys, the receiving ends might not be able receive the
> > > XOFF characters and just keep sending more data to the opposite
> > > direction. This problem is almost guaranteed to occur with DMA
> > > which sends data in large chunks.
> > >
> > > If TTY is slow to process characters, that is, eats less than given
> > > amount in receive_buf, invoke lookahead for the rest of the chars
> > > to process potential XON/XOFF characters.
> > >
> > > The guards necessary for ensuring the XON/XOFF character are
> > > processed only once were added by the previous patch. All this patch
> > > needs to do on that front is to pass the lookahead count (that can
> > > now be non-zero) into port->client_ops->receive_buf().
> >
> > ...
> >
> > > +static bool __n_tty_receive_char_special(struct tty_struct *tty, unsigned char c,
> > > + bool lookahead_done)
> > > +{
> > > + if (!I_IXON(tty))
> > > + return false;
> > > +
> > > + if (c == START_CHAR(tty)) {
> > > + if (!lookahead_done) {
> > > + start_tty(tty);
> > > + process_echoes(tty);
> > > + }
> > > + return true;
> > > + }
> > > + if (c == STOP_CHAR(tty)) {
> > > + if (!lookahead_done)
> > > + stop_tty(tty);
> > > + return true;
> > > + }
> > > + return false;
> > > +}
> >
> > Looking into this I would first make a preparatory patch that splits out
> > current code into something like
> >
> > static bool __n_tty_receive_char_special_no_lookahead(struct tty_struct *tty, unsigned char c)
> > {
> > ...current code...
> > }
> >
> > Then in the patch 1 you add
> >
> > static bool __n_tty_receive_char_special_lookahead(struct tty_struct *tty, unsigned char c)
> > {
> > ...
> > }
> >
> > static bool __n_tty_receive_char_special(struct tty_struct *tty, unsigned char c,
> > bool lookahead_done)
>
> This should be dubbed better. Maybe n_tty_receive_char_flow_control()?
>
> And I would place the if (I_IXON(tty)) to the caller. I am a bit lost in
> this pseudo code, so maybe this doesn't make sense in your proposal. I have
> something like in my mind:
>
> if (I_IXON(tty))
> return n_tty_receive_char_flow_control();

My point to have three helpers which make each change cleaner:

.-> n_tty_receive_char_flow_control_lah()
|
| .-> n_tty_receive_char_flow_control_no_lah()
| |
`- + -- n_tty_receive_char_flow_control()

Where no_lah variant can be split as preparatory patch prepending the current
series.

And yes, calling I_IXON at the caller seems better.

> Historically, these n_tty_receive* function names were a big mess. Don't
> produce more of that by simply prepending only "__".
>
> > {
> > if (!I_IXON(tty))
> > return false;
> >
> > if (lookahead_done)
> > return _lookahead();
> >
> > return _no_lookahead();
> > }

--
With Best Regards,
Andy Shevchenko