Re: [PATCH v5 1/2] ttynull: Add an option to allow ttynull to be used as a console device

[Petr Mladek <pmladek@suse.com>]

Hi Adam,

please add printk maintainers into Cc as already suggested by Andy
at https://lore.kernel.org/r/CAHp75VeBaetiQBykfLk_weBHdzZF1nWp=k8BJu+OKNp6iYRRTg@mail.gmail.com

The motivation is that the console registration code is in
kernel/printk/printk.c. It is historically pretty tricky.
Some ordering is defined rather by chance than by design.
And we should be careful when adding new rules and hacks.

On Mon 2025-02-24 07:39:14, adamsimonelli@gmail.com wrote:
> From: Adam Simonelli <adamsimonelli@gmail.com>
> 
> The new config option, CONFIG_NULL_TTY_CONSOLE will allow ttynull to be
> initialized by console_initcall() and selected as a possible console
> device.
> 
> Signed-off-by: Adam Simonelli <adamsimonelli@gmail.com>
> ---
>  drivers/tty/Kconfig   | 15 ++++++++++++++-
>  drivers/tty/ttynull.c | 20 +++++++++++++++++++-
>  2 files changed, 33 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/tty/Kconfig b/drivers/tty/Kconfig
> index 63a494d36a1f..b4afae8b0e74 100644
> --- a/drivers/tty/Kconfig
> +++ b/drivers/tty/Kconfig
> @@ -383,7 +383,20 @@ config NULL_TTY
>  	  available or desired.
>  
>  	  In order to use this driver, you should redirect the console to this
> -	  TTY, or boot the kernel with console=ttynull.
> +	  TTY, boot the kernel with console=ttynull, or enable
> +	  CONFIG_NULL_TTY_CONSOLE.
> +
> +	  If unsure, say N.
> +
> +config NULL_TTY_CONSOLE

It makes sense to enable this behavior by a CONFIG_ setting
but the name is misleading.

> +
> +        bool "Support for console on ttynull"
> +        depends on NULL_TTY=y && !VT_CONSOLE
> +	help
> +	  Say Y here if you want the NULL TTY to be used as a /dev/console
> +	  device.
> +
> +	  This is similar to CONFIG_VT_CONSOLE, but without the dependency on
> +	  CONFIG_VT. It uses the ttynull driver as the system console.

It is true that CONFIG_VT_CONSOLE causes that the virtual terminal
will get associated with /dev/console. But it works only "by chance".

It works because "register_console(&vt_console_driver)" in
con_init() is the first register_console() call. And it also
works only by chance because of the linking order.

Anyway, there are more similar CONFIG_ options, for example,
CONFIG_LP_CONSOLE, or CONFIG_VIRTIO_CONSOLE. And they are not
default when CONFIG_VT_CONSOLE is enabled. They are registered only
when the related console= option is defined on the command line.

I want to say that CONFIG_<BLA>_CONSOLE does not mean
that the BLA console will be registered by default.
And we should us a better descriptive name, for example,

  NULL_TTY_DEFAULT_CONSOLE
  NULL_TTY_DEV_CONSOLE

>  	  If unsure, say N.
>  
> diff --git a/drivers/tty/ttynull.c b/drivers/tty/ttynull.c
> index 6b2f7208b564..ec3dd3fd41c0 100644
> --- a/drivers/tty/ttynull.c
> +++ b/drivers/tty/ttynull.c
> @@ -57,6 +57,13 @@ static struct tty_driver *ttynull_device(struct console *c, int *index)
>  static struct console ttynull_console = {
>  	.name = "ttynull",
>  	.device = ttynull_device,
> +
> +	/*
> +	 * Match the index and flags from other boot consoles when CONFIG_NULL_TTY_CONSOLE is
> +	 * enabled, otherwise, use the default values for the index and flags.
> +	*/
> +	.index = IS_ENABLED(CONFIG_NULL_TTY_CONSOLE) ? -1 : 0,

This should not be needed. "con->index" is always initialized to "0"
for the default console, see:

static void try_enable_default_console(struct console *newcon)
{
	if (newcon->index < 0)
		newcon->index = 0;
[...]
}

> +	.flags = IS_ENABLED(CONFIG_NULL_TTY_CONSOLE) ? CON_PRINTBUFFER : 0,

This does not make much sense to me.

CON_PRINTBUFFER prevents duplicated output when the same device has
already been registered as a boot console. But ttynull does not have
a boot console variant. Also it is a "null" device. It never prints
anything. The output could never be duplicated by definition.

>  };
>  
>  static int __init ttynull_init(void)
> @@ -90,11 +97,22 @@ static int __init ttynull_init(void)
>  	}
>  
>  	ttynull_driver = driver;
> -	register_console(&ttynull_console);
> +	if (!console_is_registered(&ttynull_console))
> +		register_console(&ttynull_console);
>  
>  	return 0;
>  }
>  
> +#ifdef CONFIG_NULL_TTY_CONSOLE
> +static int __init ttynull_register(void)
> +{
> +	if (!console_is_registered(&ttynull_console))
> +		register_console(&ttynull_console);
> +	return 0;
> +}
> +console_initcall(ttynull_register);
> +#endif

This looks strange. I guess that you needed to move this into
console_initcall() because it is called earlier together
with the other console_initcall() calls for serial ports.
Otherwise, the hack with the linking order (2nd patch) did
not work.

But you needed to keep it in ttynull_init() so that ttynull
did not get registered prematurely when CONFIG_NULL_TTY_CONSOLE
was not enabled.

Sigh, it looks like a dirty hack which works rather by chance
than by design.


Thinking loudly:

The register_console() code is a historic mess. I dream about
having time to clean it up. Anyway, there are basically two
modes:

1. try_enable_default_console(newcon) is called only when
   there is no @preferred_console and there is no registered
   console with tty binding (valid con->device).

   The first register_console() caller wins. The order is defined
   by the __con_initcall section. Which is defined by the linking
   order.

   IMHO, it is quite fragile and non-intuitive.


2. try_enable_preferred_console() is called when some
   console is preferred via console_cmdline[]. The entries
   are added by __add_preferred_console() calls.

   This approach was created to handle console= command line
   parameters. But it was later used to define default consoles
   also via SPCR and device tree, see add_preferred_console()
   callers.

   It is also a bit tricky because the last added entry
   is preferred. Plus the .user_specified entries are
   preferred over the entries added via SPCR or device tree.

   Anyway, I think that the preference and ordering defined
   by console_cmdline[] array is a more intuitive approach.


My proposal is to call:

#ifdef CONFIG_NULL_TTY_DEFAULT_CONSOLE
	add_preferred_console("ttynull", 0, NULL);
#endif

somewhere in the kernel code. The question is where.
I wonder if the following would work:

#ifdef CONFIG_NULL_TTY_DEFAULT_CONSOLE
static int __init ttynull_default_console(void)
{
	add_preferred_console("ttynull", 0, NULL);
	return 0;
}
console_initcall(ttynull_register);
#endif

Best Regards,
Petr