On Sat 2020-12-05 13:39:53, Sergey Senozhatsky wrote:
> On (20/12/04 17:19), Petr Mladek wrote:
> [..]
> > > --- a/kernel/printk/printk.c
> > > +++ b/kernel/printk/printk.c
> > > @@ -2432,7 +2490,6 @@ void console_unlock(void)
> > > size_t len;
> > >
> > > printk_safe_enter_irqsave(flags);
> >
> > Why do we actually need to use the printk_safe context here?
> > There is not longer a risk of deadlock caused by logbuf_lock.
> > All other recursions should be prevented by console_trylock()
> > in printk().
>
> All semaphore functions, including down_trylock(), acquire
> semaphore spin_lock;
This has a very easy solution. The patch adds a code that counts
recursion level. We just must not call console_trylock() when
being inside a recursive printk.
printk_safe() has two functions:
1. It allows to store messages a lockless way. This is obsoleted
by the lockless ringbuffer.
2. It prevents calling consoles. We could move this check
into vprintk_store(). We already have there similar check
for printk_deferred().
> and then some call into the scheduler
> (or other kernel core functions) under semaphore's spin_lock.
> For instance
>
> up()
> raw_spin_lock_irqsave(&sem->lock)
> __up()
> wake_up_process()
> try_to_wake_up()
This problem is partly solved by printk_deferred(). In each
case, printk_safe() does not help here.
I still do _not_ see a reason to keep printk_safe()!
Best Regards,
Petr
