Re: [RFC][PATCHv2 1/4] panic: avoid deadlocks in re-entrant console drivers

From: Sergey Senozhatsky
Date: Thu Oct 25 2018 - 05:05:49 EST

On (10/25/18 10:29), Petr Mladek wrote:
> Yes, klogd is not a big deal. I just think that the bust_spinlocks()
> ping-pong would just confuse the code.

I agree; that's why I put some comments there.

> It might be better to keep the spinlocks busted and make sure that we do
> not cause regressions by not calling bust_spinlocks(0).

Sure, I understand. One reason to keep bust_spinlocks(0) there was "invoke
arch-specific bust_spinlocks(0), which might do something that common
bust_spinlocks() wouldn't do". Without going into details if any arch actually
does anything "special" in bust_spinlocks(0). Another reason was - this patch
looks like a -stable material to me; especially given that we have panic()
deadlock reports now. So I wanted to have a one liner which will not change
things for arch-s that re-define bust_spinlocks() and, at the same time,
fix the deadlock. Other than that I'm all for keeping spinlocks busted all
the time and just doing:

#ifdef CONFIG_VT

in panic().

BTW, speaking of s390 bust_spinlocks(). It seems that starting from 4.21
all arch-s will use common bust_spinlocks() [1].

> > Hmm, I don't think I've seen any reports because of this. From printk/console
> > POV the locks which are not taken under oops_in_progress are not released.
> Fair enough. Let's keep debug_locks_off() in panic().