From mboxrd@z Thu Jan  1 00:00:00 1970
From: tglx@linutronix.de (Thomas Gleixner)
Date: Tue, 17 May 2011 21:31:06 +0200 (CEST)
Subject: [PATCH] arm: Use arch_spin_lock for unwind to avoid lockdep
	deadlock
In-Reply-To: <alpine.LFD.2.02.1105171737260.3078@ionos>
References: <alpine.LFD.2.02.1105171605080.3078@ionos>
	<1305642021.2466.5799.camel@twins>
	<alpine.LFD.2.02.1105171737260.3078@ionos>
Message-ID: <alpine.LFD.2.02.1105172115560.3078@ionos>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On Tue, 17 May 2011, Thomas Gleixner wrote:

> On Tue, 17 May 2011, Peter Zijlstra wrote:
> 
> > On Tue, 2011-05-17 at 16:11 +0200, Thomas Gleixner wrote:
> > > When lockdep traces a lock taken in a module the arm unwinder takes
> > > unwind_lock which causes a recursive lockdep call deadlocking itself
> > > on the unwind_lock.
> > > 
> > > Make unwind_lock an arch_spin_lock to avoid the lockdep call and use
> > > the raw_local_irq_save/restore() variants to avoid irq tracing as
> > > well. 
> > 
> > That doesn't parse, lockdep has recursion protection, this shouldn't
> > happen.
> 
> Gah, misread the splat. The problem is 
> 
>      unwind() -> lock() -> lock_acquired() -> unwind() -> lock()

Grmpf, no. That does not make sense as the stack trace is taken in
lock_aquire() which happens before lock(). Ignore that for now until I
have it reproduced.

Thanks,

	tglx