Re: [PATCH next v2 2/3] printk: change @clear_seq to atomic64_t

[Peter Zijlstra]

On Wed, Dec 09, 2020 at 07:46:13PM +0900, Sergey Senozhatsky wrote:
> On (20/12/09 18:22), Sergey Senozhatsky wrote:
> > > 
> > > Please put on your eye cancer gear and inspect the atomic implementation
> > > of PA-RISC, Sparc32, feh, I forgot who else.
> > > 
> > > Those SMP capable architectures are gifted with just one XCHG like
> > > atomic instruction :/ Anyway, as said in the other email, they also
> > > don't have NMIs so it mostly works.
> 
> PeterZ, thanks for the pointers!
> 
> 
> > Hmm, wow. OK, I definitely want to look further.
> > 
> > When some CONFIG_DEBUG_FOO_BAR code wants to pr_err from prb->atomic_op
> > on those archs then we deadlock in printk once again?
> 
> E.g. arch/sparc/lib/atomic32.c
> 
> 	spinlock_t __atomic_hash[ATOMIC_HASH_SIZE];
> 	atomic_foo()
> 	{
> 		spin_lock_irqsave(ATOMIC_HASH(v), flags)
> 		...
> 		spin_unlock_irqrestore(ATOMIC_HASH(v), flags);
> 	}
> 
> So another potential re-entry path is
> 
> 	atomic_foo()
> 	 spin_lock_irqsave(ATOMIC_HASH(v), flags)
> 	  printk()
> 	   prb()
> 	    atomic_foo()
> 	     spin_lock_irqsave(ATOMIC_HASH(v), flags)
> 
> which can deadlock, in theory, if both atomics HASH to the same
> key (same spin_lock).

Yep, but see the 'mostly' in the 'they mostly work'. Given the
limitiations of these architectures there's really only so much you can
do.