2.6.13-rc1 CONFIG_DEBUG_SPINLOCK is useless on SMP

2.6.13-rc1 CONFIG_DEBUG_SPINLOCK is useless on SMP

[Keith Owens]

2.6.13-rc1 built with SMP=Y and DEBUG_SPINLOCK=y.  That uses
kernel/spinlock.c instead of the inline definitions of the spinlock
functions.  Alas only the inline definitions test for DEBUG_SPINLOCK=y,
none of the code in in spinlock.c has any debug facilities.

Re: 2.6.13-rc1 CONFIG_DEBUG_SPINLOCK is useless on SMP

[Ingo Molnar]

* Keith Owens <kaos@ocs.com.au> wrote:

> 2.6.13-rc1 built with SMP=Y and DEBUG_SPINLOCK=y.  That uses 
> kernel/spinlock.c instead of the inline definitions of the spinlock 
> functions.  Alas only the inline definitions test for 
> DEBUG_SPINLOCK=y, none of the code in in spinlock.c has any debug 
> facilities.

the spinlock cleanups in -mm (not yet in -git) obsolete most of the 
debugging approach of the current kernel.

but even in terms of 2.6.13-rc1, i'm not completely sure what you mean.
Yes, there's no debugging code in kernel/spinlock.c because much of the
DEBUG_SPINLOCK code (under the old method) is located in the arch
spinlock.h files. I.e. on x86 you'll get SMP spinlock debugging from
asm-i386/spinlock.h:

 static inline void _raw_spin_lock(spinlock_t *lock)
 {
 #ifdef CONFIG_DEBUG_SPINLOCK
         if (unlikely(lock->magic != SPINLOCK_MAGIC)) {
                 printk("eip: %p\n", __builtin_return_address(0));
                 BUG();
         }
 #endif

the only practical exception is when CONFIG_PREEMPT is enabled: the arch 
level trylock, upon which the PREEMPT spinlocks rely on heavily, has no 
meaningful DEBUG_SPINLOCK checks. (but PREEMPT has other checks which 
partly offset this.) In any case, this too is fixed by my spinlock 
cleanups - there all debugging is done centrally in 
lib/spinlock_debug.c.

	Ingo