Re: [PATCHv7 6/8] printk: use printk_safe buffers in printk

[Petr Mladek <pmladek@suse.com>]

On Wed 2017-02-01 10:37:39, Peter Zijlstra wrote:
> On Wed, Feb 01, 2017 at 10:06:25AM +0100, Jan Kara wrote:
> > Clearly scheduler code (update_load_avg) calls WARN_ON from scheduler while
> > holding rq_lock which has been always forbidden. Sergey and Petr were doing
> > some work to prevent similar deadlocks but I'm not sure how far they
> > went...
> 
> Its not forbidden, just can result in the occasional deadlock, meh.
> 
> 
> In any case, there's a patch in tip fixing that warn trigger.

I guess that you are talking about the introduction of
#define SCHED_WARN_ON(x)	WARN_ONCE(x, #x)

It reduces the risk of the deadlock but some risk is still there.
IMHO, it does not avoid the lockdep warning.

One solution would be to hide the occasional deadlock and disable
lockdep in SCHED_WARN_ON():

#define SCHED_WARN_ON(x)				\
({							\
	int __ret_sched_warn_on;			\
	lockdep_off();					\
	__ret_sched_warn_on = WARN_ONCE(x, #x);		\
	lockdep_on();					\
	unlikely(__ret_sched_warn_on);			\
})


Another solution would be to redirect it into the
alternative buffer and let it printed later:

#define SCHED_WARN_ON(x)	WARN_ONCE(x, #x)		\
({								\
	unsigned long __sched_warn_on_flags;			\
	printk_safe_enter_irqsave(__sched_warn_on_flags);	\
	__ret_sched_warn_on = WARN_ONCE(x, #x);			\
	printk_safe_exit_irqrestore(__sched_warn_on_flags);	\
	unlikely(__ret_sched_warn_on);				\
})

Best Regards,
Petr