Re: [PATCH v3 5/5] console: Introduce ->exit() callback

[Petr Mladek]

On Wed 2020-01-29 16:25:58, Andy Shevchenko wrote:
> On Wed, Jan 29, 2020 at 10:41:41PM +0900, Sergey Senozhatsky wrote:
> > On (20/01/28 11:44), Andy Shevchenko wrote:
> > > > If the console was not registered (hence not enabled) is it still required
> > > > to call ->exit()? Is there a requirement that ->exit() should handle such
> > > > cases?
> > > 
> > > This is a good point. The ->exit() purpose is to keep balance for whatever
> > > happened at ->setup().
> > > 
> > > But ->setup() is being called either when we have has_preferred == false or
> > > when we got no matching we call it for all such consoles, till it returns an
> > > error (can you elaborate the logic behind it?).
> > 
> > ->match() does alias matching and ->setup(). If alias matching failed,
> > exact name match takes place. We don't call ->setup() for all consoles,
> > but only for those that have exact name match:
> > 
> > 	if (strcmp(c->name, newcon->name) != 0)
> > 		continue;
> > 
> > As to why we don't stop sooner in that loop - I need to to do some
> > archaeology. We need to have CON_CONSDEV at proper place, which is
> > IIRC the last matching console.
> > 
> > Pretty much every time we tried to change the logic we ended up
> > reverting the changes.
> 
> I understand. Seems the ->setup() has to be idempotent. We can tell the same
> for ->exit() in some comment.

I believe that ->setup() can succeesfully be called only once.
It is tricky like hell:

1st piece:

	if (!has_preferred || bcon || !console_drivers)
		has_preferred = preferred_console >= 0;

  note:

     + "has_preferred" is updated here only when it was not "true" before.
     + "has_preferred" is set to "true" here only when "preferred_console"
       is set in __add_preferred_console()

2nd piece:

  + __add_preferred_console() is called for console defined on
    the command line. "preferred_console" points to the console
    defined by the last "console=" parameter.

3rd piece:

  + "has_preferred" is set to "true" later in register_console() when
    a console with tty binding gets enabled.

4th piece:

  + The code:

	/*
	 *	See if we want to use this console driver. If we
	 *	didn't select a console we take the first one
	 *	that registers here.
	 */
	if (!has_preferred)
		... try to enable the given console

   The comment is a bit unclear. The code is used as a fallback
   when no console was defined on the command line.

   Note that "has_preferred" is always true when "preferred_console"
   was defined via command line, see 2nd piece above.


By other words:

  + The fallback code (4th piece) is called only when
    "preferred_console" was not defined on the command line.

  + The cycle below matches the given console only when
    it was defined on the command line.


As a result, I believe that ->setup() could never be called
in both paths for the same console. Especially I think that
fallback code should not be used when the console was defined on
the command line.

I am not 100% sure but I am ready to risk this. Anyway, I think
that many ->setup() callbacks are not ready to be successfully
called twice.

(Sigh, I have started to clean up this code two years ago.
But I have never found time to finish the patchset. It is
such a huge mess.)

> Can you describe, btw, struct console in kernel doc format?
> It will be very helpful!
> 
> > > In both cases we will get the console to have CON_ENABLED flag set.
> > 
> > And there are sneaky consoles that have CON_ENABLED before we even
> > register them.
> 
> So, taking into consideration my comment to the previous patch, what would be
> suggested guard here?
> 
> For a starter something like this?
> 
>   if ((console->flags & CON_ENABLED) && console->exit)
> 	console->exit(console);

I would do:

	if (!res && console->exit)
		console->exit(console);

I mean. I would call ->exit() only when console->setup() succeeded in
register_console(). In this case, the console was later added to
the console_drivers list.

Best Regards,
Petr