Re: [RFC][PATCH 6/7] printk: use alternative printk buffers

[Petr Mladek <pmladek@suse.com>]

On Fri 2016-09-30 10:15:44, Sergey Senozhatsky wrote:
> On (09/29/16 15:00), Petr Mladek wrote:
> [..]
> > > @@ -1791,7 +1791,7 @@ asmlinkage int vprintk_emit(int facility, int level,
> > >  		zap_locks();
> > >  	}
> > >  
> > > -	lockdep_off();
> > > +	alt_printk_enter();
> > 
> > IMHO, we could not longer enter vprintk_emit() recursively. The same
> > section that was guarded by logbuf_cpu is guarded by
> > alt_printk_enter()/exit() now.
> 
> you might be very right here. I'll take a look.
> 
> > IMHO, we could remove all the logic around the recursion. Then we
> > could even disable/enable irqs inside alt_printk_enter()/exit().
> 
> I was thinking of doing something like this; but that would require
> storing 'unsigned long' flags in per-cpu data
> 
> 	alt_enter()
> 	{
> 		unsinged long flags;
> 
> 		local_irq_save(flags);
> 		ctx = this_cpu_ptr();
> 		ctx->flags = flags;
> 		...
> 	}
> 
> 	alt_exit()
> 	{
> 		ctx = this_cpu_ptr();
> 		...
> 		local_irq_restore(ctx->flags);
> 	}
>
> and the decision was to keep `unsigned long flags' on stack in the
> alt_enter/exit caller. besides in most of the cases we already have
> it (in vprintk_emit() and console_unlock()).

I would pass the pointer to flags as alt_enter() parameter.
> 
> but I can certainly hide these details in alt_enter/exit.
>
> > > @@ -2479,7 +2490,9 @@ void console_unlock(void)
> > >  	 */
> > >  	raw_spin_lock(&logbuf_lock);
> > >  	retry = console_seq != log_next_seq;
> > > -	raw_spin_unlock_irqrestore(&logbuf_lock, flags);
> > > +	raw_spin_unlock(&logbuf_lock);
> > > +	alt_printk_exit();
> > > +	local_irq_restore(flags);
> > 
> > We should mention that this patch makes an obsolete artefact from
> > printk_deferred(). It opens the door for another big cleanup and
> > relief.
> 
> do you mean that, once alt_printk is done properly, we can drop
> printk_deferred()? I was thinking of it, but decided not to
> mention/touch it in this patch set.

My understanding is the following:

The difference between normal printk() and printk_deferred() is
that the other does not call console_trylock()/console_unlock().
It means that printk_deferred() can avoid recursion only from these
two calls.

printk_deferred() is used only in scheduler and timekeeping code.
Therefore it prevents only limited number of possible recursions
and deadlocks at the moment.

This patch guards most of the two calls a more generic way.
The redirected parts prevent recursion not only to into the
code guarded by console_sem but also into parts guarded
by lockbuf_lock.

By other words, this patch is supposed to handle a superset
of the deadlocks that are currently prevented by printk_deferred().
If this is true, we do not longer need printk_deferred().

The only question is if this patch guards enough parts of
console_try_lock()/console_unlock() to handle the superset
of the possible deadlocks.

I see that it does not guard two up_console_sem() calls
from console_unlock(). But this can be fixed in the next
version.

Or is there any other catch that I do not see at the moment?

In each case, getting rid of printk_deferred() could be
a fantastic selling point for this patchset.

Best Regards,
Petr