functionality: was: Re: [patch RFC 19/29] printk: Add basic infrastructure for non-BKL consoles

[Petr Mladek <pmladek@suse.com>]

On Sun 2022-09-11 00:28:01, Thomas Gleixner wrote:
> The current console/printk subsystem is protected by a Big Kernel Lock,
> aka. console_lock which has has ill defined semantics and is more or less
> stateless. This puts severe limitations on the console subsystem and makes
> forced takeover and output in emergency and panic situations a fragile
> endavour which is based on try and pray.
> 
> The goal of non-BKL consoles is to break out of the console lock jail and
> to provide a new infrastructure which avoids the pitfalls and allows
> console drivers to be gradually converted over.
> 
> The proposed infrastructure aims for the following properties:
> 
>   - Lockless (SCRU protected) console list walk
>   - Per console locking instead of global locking
>   - Per console state which allows to make informed decisions
>   - Stateful handover and takeover
> 
> As a first step this adds state to struct console. The per console state is
> a atomic_long_t with a 32bit bit field and on 64bit a 32bit sequence for
> tracking the last printed ringbuffer sequence number. On 32bit the sequence
> is seperate from state for obvious reasons which requires to handle a few
> extra race conditions.
> 
> Add the initial state with the most basic 'alive' and 'enabled' bits and
> wire it up into the console register/unregister functionality and exclude
> such consoles from being handled in the console BKL mechanisms.
> 
> The decision to use a bitfield was made as using a plain u32 and mask/shift
> operations turned out to result in uncomprehensible code.
> 
> --- a/include/linux/console.h
> +++ b/include/linux/console.h
> @@ -170,6 +172,37 @@ enum cons_flags {
>  	CON_ANYTIME		= BIT(4),
>  	CON_BRL			= BIT(5),
>  	CON_EXTENDED		= BIT(6),
> +	CON_NO_BKL		= BIT(7),
> +};
> +
> +/**
> + * struct cons_state - console state for NOBKL consoles
> + * @atom:	Compound of the state fields for atomic operations
> + * @seq:	Sequence for record tracking (64bit only)
> + * @bits:	Compound of the state bits below
> + *
> + * @alive:	Console is alive. Required for teardown

What do you exactly mean with teardown, please?

I somehow do not understand the meaning. The bit "alive" seems
to always be "1" in this patchset.

> + * @enabled:	Console is enabled. If 0, do not use
> + *
> + * To be used for state read and preparation of atomic_long_cmpxchg()
> + * operations.
> + */
> +struct cons_state {
> +	union {
> +		unsigned long	atom;
> +		struct {
> +#ifdef CONFIG_64BIT
> +			u32	seq;
> +#endif
> +			union {
> +				u32	bits;
> +				struct {
> +					u32 alive	:  1;
> +					u32 enabled	:  1;
> +				};
> +			};
> +		};
> +	};
>  };
>  
>  /**
> --- a/kernel/printk/printk.c
> +++ b/kernel/printk/printk.c
> @@ -3079,7 +3088,10 @@ void console_stop(struct console *consol
>  	console_list_lock();
>  	console_lock();
>  	console->flags &= ~CON_ENABLED;
> +	cons_state_disable(console);
>  	console_unlock();
> +	/* Ensure that all SRCU list walks have completed */
> +	synchronize_srcu(&console_srcu);

I have few questions here:

1. Do we need separate "enabled" flags for BLK and non-blk consoles?

   Hmm, it might be problem to remove CON_ENABLED flag
   because it is exported to userspace via /proc/consoles.

   Well, what is the purpose of the "enabled" flag for atomic
   consoles? Are we going to stop them in the middle of a line?
   Does the flag has to be atomic and part of atomic_state?


2. What is the purpose of synchronize_srcu(), please?

   It probably should make sure that all consoles with CON_NO_BLK
   flag are really stopped once it returns.

   IMHO, this would work only when the "enabled" flag and the
   con->write*() callback is called under srcu_read_lock().

   I do not see it in the code. Do I miss something, please?


3. Is the ordering of console_unlock() and synchronize_srcu()
   important, please?

   IMHO, it would be important if we allowed the following code:

      srcu_read_lock(&console_srcu);
      console_lock();
      // do something
      console_unlock();
      srcu_read_unlock(&console_srcu);

   then we would always have to call synchronize_srcu() outside
   console_lock() otherwise there might be ABBA deadlock.

   I do not see this code yet. But it might make sense.
   Anyway, we should probably document the rules somewhere.


4. Is it important to call cons_state_disable(console) under
   console_lock() ?

   I guess that it isn't. But it is not clear from the code.
   The picture is even more complicated because everything is done
   under console_list_lock().

   It would make sense to explain the purpose of each lock.
   My understanding is the following:

      + console_list_lock() synchronizes manipulation of
	con->flags.

      + console_lock() makes sure that no console will
	be calling con->write() callback after console_unlock().

      + synchronize_srcu() is supposed to make sure that
	any console is calling neither con->write_kthread()
	nor con->atomic_write() after this synchronization.
	Except that it does not work from my POV.

    Anyway, I might make sense to separate the two approaches.
    Let's say:

	console_list_lock()
	if (con->flags & CON_NO_BLK) {
		noblk_console_disable(con);
	} else {
		/* cons->flags are synchronized using console_list_lock */
		console->flags &= ~CON_ENABLED;
		/*
		 * Make sure that no console calls con->write()	anymore.
		 *
		 * This ordering looks a bit ugly. But it shows how
		 * the things are serialized.
		 */
		console_lock();
		console_unlock();
	}

     , where noblk_console_disable(con) must be more complicated.
     It must be somehow synchronized with all con->write_kthread() and
     write_atomic() callers.

     I wonder if noblk_console_disable(con) might somehow use
     the hangover mechanism so that it becomes the owner of
     the console and disables the enabled flag. I mean
     to implement some sleepable cons_acquire(). But this sounds
     a bit like con->mutex that you wanted to avoid.

     It might be easier to check the flag and call con->write() under
     srcu_read_lock() so that synchronize_srcu() really waits until
     the current message gets printed.


>  	console_list_unlock();
>  }
>  EXPORT_SYMBOL(console_stop);


Best Regards,
Petr

PS: I am going to review v3 of "reduce console_lock scope" patchset
    which has arrived few hours ago.

    I just wanted to send my notes that I made last Friday
    when I continued review of this RFC.