Re: [patch RFC 06/29] printk: Protect [un]register_console() with a mutex

[Petr Mladek <pmladek@suse.com>]

On Sun 2022-09-11 00:27:41, Thomas Gleixner wrote:
> Unprotected list walks are a brilliant idea. Especially in the context of
> hotpluggable consoles.

Yeah, it is crazy. And it is there probably since the beginning.

> --- a/kernel/printk/printk.c
> +++ b/kernel/printk/printk.c
> @@ -79,10 +79,14 @@ int oops_in_progress;
>  EXPORT_SYMBOL(oops_in_progress);
>  
>  /*
> - * console_sem protects the console_drivers list, and also
> - * provides serialisation for access to the entire console
> - * driver system.
> + * console_sem protects the console_drivers list, and also provides
> + * serialization for access to the entire console driver system.
> + *
> + * console_mutex serializes register/unregister. console_sem has to be
> + * taken for any list manipulation inside the console_mutex locked
> + * section to keep the console BKL machinery happy.
>   */
> +static DEFINE_MUTEX(console_mutex);
>  static DEFINE_SEMAPHORE(console_sem);
>  struct console *console_drivers;
>  EXPORT_SYMBOL_GPL(console_drivers);
> @@ -220,6 +230,26 @@ int devkmsg_sysctl_set_loglvl(struct ctl
>  }
>  #endif /* CONFIG_PRINTK && CONFIG_SYSCTL */
>  
> +/**
> + * console_list_lock - Lock the console list
> + *
> + * For non-console related list walks, e.g. procfs, sysfs...
> + */
> +void console_list_lock(void)
> +{
> +	mutex_lock(&console_mutex);
> +}
> +
> +/**
> + * console_list_unlock - Unlock the console list
> + *
> + * Counterpart to console_list_lock()
> + */
> +void console_list_unlock(void)
> +{
> +	mutex_unlock(&console_mutex);
> +}
> +
>  /*
>   * Helper macros to handle lockdep when locking/unlocking console_sem. We use
>   * macros instead of functions so that _RET_IP_ contains useful information.
> @@ -3107,13 +3143,14 @@ void register_console(struct console *ne
>  	bool realcon_enabled = false;
>  	int err;
>  
> -	for_each_console(con) {
> +	console_list_lock();

Hmm, the new mutex is really nasty. It has very strange semantic.
It makes the locking even more complicated.

The ideal solution would be take console_lock() here. We (me and
Sergey) never did it because con->match() and con->setup()
callbacks were called in try_enable_*console(). We were afraid
that some might want to take console_lock() and it could create
a deadlock. There were too many drivers and we did not found time
to check them all. And it had low priority because nobody reported
problems.

A good enough solution might be call this under the later
added srcu_read_lock(&console_srcu) and use for_each_console_srcu().

The srcu walk would prevent seeing broken list. Obviously,
the code might see outdated list and do bad decisions:

  + try to enable the same console twice

  + enable more consoles by default in try_enable_default_console()

  + associate more consoles with /dev/console, see CON_CONSDEV in
    try_enable_preferred_console() and try_enable_default_console()

If we race then we could end up with more consoles enabled by default
and with more consoles with CON_CONSDEV flag.

IMHO, the rcu walk is an acceptable and conservative solution.
Registering the same driver twice is hard to imagine at all.
And I have never seen reports about too many default consoles
or CON_CONSDEV flags.

Anyway, I would like to avoid adding console_mutex. From my POV,
it is a hack that complicates the code. Taking console_lock()
should be enough. Using rcu walk would be good enough.

Do I miss something, please?

> +	for_each_registered_console(con) {
>  		if (WARN(con == newcon, "console '%s%d' already registered\n",
>  					 con->name, con->index))
> -			return;
> +			goto unlock;
>  	}
>  
> -	for_each_console(con) {
> +	for_each_registered_console(con) {
>  		if (con->flags & CON_BOOT)
>  			bootcon_enabled = true;
>  		else

Best Regards,
Petr