Re: [PATCH printk v2 2/5] printk: Add NMI safety to console_flush_on_panic() and console_unblank()

[John Ogness]

On 2023-07-11, Petr Mladek <pmladek@xxxxxxxx> wrote:
> Just to be sure. The semaphore is not NMI safe because even the
> trylock takes an internal spin lock. Am I right, please?

Yes, that is one of the reasons. Sergey mentioned another (waking a task
on up()).

> Alternative solution would be to make down_trylock() NMI safe
> by using raw_spin_trylock_irqsave() for the internal lock.

NMI contexts are only allowed to take raw spinlocks if those spinlocks
are only used from NMI context. Otherwise you could have deadlock:

raw_spin_lock()
--- NMI ---
raw_spin_lock()

Using a trylock does not avoid the deadlock danger.

> Another question is whether we want to call c->unblank()
> in NMI even when down_trylock() was NMI safe. It seems that it
> is implemented only for struct console vt_console_driver.
> I am pretty sure that it takes more internal locks which
> are not NMI safe either.

Yes, it does. As an example, it calls mod_timer(), which is also not NMI
safe. Clearly the unblank() callback must not be called in NMI context.

> Finally, it is not only about NMI. Any locks might cause a deadlock
> in panic() in any context. It is because other CPUs are stopped
> and might block some locks.

With the atomic/threaded model this is not true. The port ownership can
be safely taken over from stopped CPUs.

> In my opinion, we should handle c->unblank() in panic() the same way
> as c->write() in panic().

I do not agree. Clearly unblank() is not NMI safe. Also, in current
mainline code, console_unblank() will already give up if the trylock
failed (rather than ignoring the lock, like write() does). So
console_unblank() might as well also give up if in NMI context.

John