Re: [RFC 0/5] printk: Implement WARN_*DEFERRED()

[Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>]

On (09/29/16 13:28), Petr Mladek wrote:
> On Wed 2016-09-28 10:18:45, Sergey Senozhatsky wrote:
> > On (09/27/16 18:02), Petr Mladek wrote:
> > > The main trick is that we replace the per-CPU function pointer
> > > by a preempt_count-like variable that could track the printk context.
> > > 
> > > I know that Sergey has another ideas in this area. But I wanted to see
> > > how this approach would look like.
> > 
> > well, yes. I was looking at WARN_*_DEFERRED [1] for some time, and, I
> > think, the maintenance cost of that solution is just too high:
> > 
> > a) every existing WARN_* in sched/timekeeping/who knows where else
> >    must be evaluated to ensure that in can't be called from printk()
> >    path. if `false' - then the corresponding macro must be replaced
> >    with _DEFERRED flavor.
> > 
> > b) any patch that adds new WARN_* usages must be additionally checked
> >    to ensure that each of new WARN_* macros cannot be called from printk
> >    path. if `false' -- the corresponding macro must be replaced with
> >    _DEFERRED flavor.
> > 
> > c) any patch that refactors the code or moves some function calls around
> >    etc. must be additionally checked for any accidental WARN_* from printk
> >    path. even though if none of the patches added any new WARN_* to the code.
> > 
> > b) apart from WARN_* there can be `accidental' pr_err/pr_debug/etc. not
> >    necessarily newly added (see 'c').
> > 
> > 
> > that's too much.
> > 
> > it takes a lot of additional effort, because both reviewer and contributor
> > must consider printk() internals. and, what's worse, if something goes
> > unnoticed we end up having a printk() deadlock again.
> > 
> > so I decided to address some of printk() issues in printk.c, not in
> > kernel/time/timekeeping.c or kernel/sched/core.c or anywhere else.
> 
> I see the point.

well, just my 5 cents.

> Your approach (alt buffer) adds some complexity to the printk code

it does.
the other thing is that there are several ways to deadlock printk().
alt_printk is addressing deadlocks that were caused by printk()
recursion only.

   printk()
     acquire_lock(&foo)
       printk()
         acquire_lock(&foo)

which is a sub-set of all of the printk() deadlock scenarios. all of
the locks that printk() acquires can be taken outside of printk() path.

for example, cat /proc/console locks the console_lock() for seq output.
thus we can have something like

        console_unlock()	// lock  &sem->lock
          up()
            activate_task()
              WARN_ON()
                printk()
                  console_trylock() // lock &sem->lock


DEFERRED_WARN is a good thing; it's just quite hard to keep everything
working, given that any of those "9 patches per hour" can break something
with just one WARN_ON().


I assume that doing something like this

#define WARN_ON(condition, format...) ({	\
	printk_deferred_enter();		\
	WARN(condition, ##format);		\
	printk_deferred_exit();			\
})

is less than exciting because WARN_ON from irq won't immediately print
the backtrace anymore.

thoughts?

> but it allows to remove printk_deferred()/WARN_DEFERRED() and all
> the risk of it.

at some point we even can drop the entire deferred_printk() thing.
but alt_printk needs some love and care first.

> I am going to look closely on it.

thanks.

	-ss